Compound proposal system, compound proposal method, and compound proposal program

Abstract

To propose two or more compounds from a group of compounds, taking into consideration the user's perception. [Solution] The compound proposal system performs regression analysis on a dataset containing multiple data records showing the compound structure and properties of a sample compound to generate a regression model that includes a model function showing the relationship between the compound structure and properties, uncertainty of the predicted values ​​by the model function, and an acquisition function based on the model function, the uncertainty, and penalties corresponding to the candidate compounds. The system then performs a search process using the regression model for each compound structure stored in the database to select two or more candidate compounds as proposed compounds. The search process includes repeatedly performing the steps of calculating the distance between the compound structure of the selected proposed compound and the compound structure of each candidate compound using an intermediate layer of a similarity determination model, and updating the penalty for each candidate compound based on that distance.

Description

Technical Field

[0001] One aspect of the present disclosure relates to a compound proposal system, a compound proposal method, and a compound proposal program.

Background Art

[0002] A mechanism for proposing specific compounds from a set of compounds using a computational model is known. For example, in Patent Document 1, a Bayesian statistical model is trained using a training set of compounds, and a step of outputting a probability distribution that approximates the characteristics of the compounds in the population as an objective function of the structural characteristics of the compounds in the population of a plurality of compounds, a step of determining a subset of a plurality of compounds from a population not included in the training set, and a step of selecting at least a part of the compounds in the determined subset are described. The subset is determined based on the probability distribution from the trained Bayesian statistical model and according to the optimization of an acquisition function based on a plurality of defined objectives.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] A mechanism for proposing two or more compounds from a plurality of compounds in a manner that also takes into account the user's sense is desired.

Means for Solving the Problems

[0005] A compound proposal system relating to one aspect of this disclosure comprises at least one processor. The at least one processor acquires a dataset containing multiple data records showing the compound structure and properties of a sample compound, performs regression analysis on the dataset to generate a regression model that includes a model function showing the relationship between compound structure and properties, uncertainty of the predicted values ​​by the model function, and an acquisition function based on the model function, the uncertainty, and penalties corresponding to the candidate compounds, and performs a search process using the regression model on the compound structure of each of the multiple candidate compounds stored in the database to select two or more candidate compounds from the multiple candidate compounds as proposed compounds. The search process includes repeatedly the steps of: selecting one candidate compound from the multiple candidate compounds as a proposed compound based on the acquisition function; calculating the distance between the compound structure of the selected proposed compound and each of the compound structures of the multiple candidate compounds using an intermediate layer of a similarity determination model that accepts input of a first compound structure and a second compound structure and determines the similarity between the first compound structure and the second compound structure; and updating the penalty corresponding to each of the multiple candidate compounds based on the distance corresponding to the candidate compound.

[0006] In this aspect, a regression model is generated that shows the relationship between the compound structure and properties of the compounds, and a search process using this regression model selects two or more candidate compounds as proposed compounds from among multiple candidate compounds. The regression model includes a model function that shows the relationship between the compound structure and properties, an acquisition function based on the uncertainty of the predicted values ​​by the model function, and penalties corresponding to the candidate compounds. In the search process, each time a proposed compound is selected, the distance between the proposed compound and each candidate compound is calculated using an intermediate layer of a similarity determination model that determines the similarity between compound structures, and the penalty for each candidate compound is updated based on these distances. Therefore, when selecting the next proposed compound, the acquisition function based on the updated penalties is used. By reflecting the distance between compound structures obtained based on the intermediate layer of the similarity determination model in the acquisition function that directly contributes to the selection of proposed compounds, it is possible to propose two or more candidate compounds from among multiple candidate compounds in a manner that also takes into account the user's perception. [Effects of the Invention]

[0007] According to one aspect of this disclosure, it is possible to propose two or more candidate compounds from a group of candidate compounds in a manner that also takes into account the user's perception. [Brief explanation of the drawing]

[0008] [Figure 1] This figure shows an example of the functional configuration of a compound proposal system. [Figure 2] This figure shows an example of the structure of a similarity detection model. [Figure 3] This flowchart shows an example of how the compound suggestion system works. [Figure 4] This figure shows an example of a search process. [Modes for carrying out the invention]

[0009] The following describes various examples in this disclosure in detail with reference to the attached drawings. In the description of the drawings, identical or equivalent elements are denoted by the same reference numeral, and redundant descriptions are omitted.

[0010] [System Overview] The compound proposal system described herein is a computer system that uses a computational model to select two or more specific compounds as proposed compounds from a set of candidate compounds. In one example, the compound proposal system presents the selected proposed compounds to the user. The user can then experiment with the proposed compounds and obtain their properties. Candidate compounds refer to compounds that may be proposed by the compound proposal system. The compound proposal system presents two or more compounds as proposed compounds from the set of candidate compounds in a manner that also takes into account the user's intuition. Simply using conventional computational algorithms such as Bayesian optimization may result in the system proposing only similar compounds that do not match the user's intuition towards the end of the search. The compound proposal system utilizes a regression model, which is a pre-trained model that includes a model function showing the relationship between compound structure and properties, and a similarity determination model, which is a pre-trained model that determines the similarity of compound structures between candidate compounds in a manner that aligns with the user's intuition, as its computational model. By using the regression model and the similarity determination model, the compound proposal system can present two or more proposed compounds from multiple candidate compounds in a manner that also takes into account the user's intuition. As a result, the user can experiment with a variety of compounds.

[0011] [System Configuration] A compound proposal system consists of one or more computers. When multiple computers are used, they are connected via a communication network such as the internet or an intranet to logically construct a single compound proposal system.

[0012] A computer comprising a compound proposal system generally includes a processor, storage device (memory), and communication interface as hardware components. The processor is, for example, a CPU, and the storage device consists of flash memory, a hard disk, etc. The communication interface consists of a network card, a wireless communication module, etc. Each functional module of the compound proposal system is realized by the processor executing a program stored in the storage device.

[0013] A compound proposal program for enabling a computer to function as a compound proposal system includes program code for implementing each functional module of the compound proposal system. This compound proposal program may be provided on a non-temporary recording medium such as a CD-ROM, DVD-ROM, or semiconductor memory. Alternatively, the compound proposal program may be provided via a communication network as a data signal superimposed on a carrier wave. The provided compound proposal program is then recorded, for example, on a storage device.

[0014] Figure 1 shows the functional configuration of a compound suggestion system 10 in one example. In this example, the compound suggestion system 10 connects to a learning database 31, a compound database 32, and a user terminal 40 via a communication network. The communication network is typically constructed by the internet, an intranet, or a combination thereof. The communication network can be constructed by a wired network, a wireless network, or a combination thereof.

[0015] The learning database 31 is a storage device that stores teacher data for generating the similarity determination model 21. The teacher data is composed of a plurality of data records (for example, a large number of data records). Each data record includes a first compound structure that is the compound structure of the first compound, a second compound structure that is the compound structure of the second compound, and the similarity between the first compound structure and the second compound structure. That is, each data record of the teacher data shows the correspondence between a pair of compound structures and the similarity. Each compound structure may be expressed in any form such as a graph representation of a chemical structure, a character string representation by SMILES (simplified molecular input line entry system), etc. The similarity is set manually and is used as the ground truth in machine learning.

[0016] The compound database 32 is a storage device that stores compound data regarding candidate compounds. Each data record of the compound data includes the compound structure of the candidate compound. Similar to the case of the teacher data, the compound structure may be expressed in any form such as a graph representation of a chemical structure, a character string representation by SMILES, etc. For example, the compound database 32 stores compound data regarding a huge number (for example, thousands, tens of thousands, etc.) of candidate compounds.

[0017] The learning database 31 and the compound database 32 may be components of the compound proposal system 10 or may be provided outside the compound proposal system 10.

[0018] The user terminal 40 is a computer used by the user of the compound proposal system 10. The user terminal 40 can be various computers such as a personal computer, a workstation, a tablet terminal, a smartphone, a wearable terminal, etc.

[0019] The compound proposal system 10 includes a processor 101 that functions as a first learning unit 11, an initial data acquisition unit 12, a second learning unit 13, a search unit 14, and a recommendation unit 15.

[0020] The first learning unit 11 is a functional module that generates a similarity determination model 21 for determining the similarity between the compound structures of two candidate compounds. The similarity between compound structures refers to an index indicating how similar two compound structures are.

[0021] The initial data acquisition unit 12 is a functional module that acquires the initial data necessary for selecting two or more proposed compounds. The initial data acquisition unit 12 receives the initial data from the user terminal 40. The initial data includes a data set for generating a regression model including a model function showing the relationship between the compound structure and properties of the sample compound, and the number of proposed compounds to be selected, which is the number of proposals. A sample compound refers to a compound shown by the data set and whose properties are known through experiments or the like.

[0022] The second learning unit 13 is a functional module that performs a regression analysis on the data set to generate a regression model 22. The regression model 22 includes a model function showing the relationship between the compound structure and properties (properties of the compound), the uncertainty of the predicted value by the model function, and an acquisition function based on the model function, the uncertainty, and a penalty corresponding to the candidate compound. The model function processes the compound structure of the candidate compound as an explanatory variable and outputs the predicted value regarding the properties of the candidate compound as the objective variable.

[0023] The search unit 14 is a functional module that performs a search process using the regression model 22 on multiple candidate compounds indicated by compound data (i.e., multiple candidate compounds stored in the compound database 32) and selects two or more candidate compounds from among the multiple candidate compounds as proposed compounds. The search unit 14 includes a penalty setting unit 14a, a compound selection unit 14b, and a distance calculation unit 14c. The penalty setting unit 14a is a functional module that sets penalties corresponding to candidate compounds. The compound selection unit 14b is a functional module that selects one candidate compound from multiple candidate compounds as a proposed compound based on an acquisition function that reflects the set penalties. The distance calculation unit 14c is a functional module that uses the intermediate layer of the similarity determination model 21 to calculate the distance between the compound structure of the selected proposed compound and the compound structures of each of the multiple candidate compounds. The search process includes the repetition of a series of processes performed in cooperation with the penalty setting unit 14a, the compound selection unit 14b, and the distance calculation unit 14c, and the search unit 14 selects two or more proposed compounds through this repetition.

[0024] The recommendation unit 15 is a functional module that presents the user with two or more selected suggested compounds. The recommendation unit 15 transmits recommendation information indicating those suggested compounds to the user terminal 40.

[0025] [System operation] The following describes the operation of the compound proposal system 10, as well as the compound proposal method related to this disclosure.

[0026] (Generation of a similarity detection model) The first learning unit 11 generates a similarity determination model 21 by performing machine learning using training data in the learning database 31. In one example, the first learning unit 11 uses a neural network model as a machine learning model for generating the similarity determination model 21, which accepts input of a first compound structure and a second compound structure and determines the similarity between the first and second compound structures. As the architecture of the neural network, a multilayer perceptron, a graph neural network, a natural language processing-based method, etc., may be employed.

[0027] The first learning unit 11 inputs the first and second compound structures, indicated by data records of training data in the learning database 31, into its machine learning model and obtains the similarity score estimated by the machine learning model. This similarity score can be expressed as the probability that the two compound structures are similar and the probability that the two compound structures are not similar. The first learning unit 11 updates the parameter set in the machine learning model by backpropagation based on the error between the estimated similarity score and the similarity score (ground truth) indicated by the data record. The first learning unit 11 repeats this process using multiple data records of the training data to generate a similarity determination model 21. The generated similarity determination model 21 is a neural network model that accepts the input of the first and second compound structures and determines the similarity between the first and second compound structures. The first learning unit 11 stores this similarity determination model 21 in a predetermined storage unit.

[0028] Figure 2 shows an example of the structure of the similarity determination model 21. In this example, the similarity determination model 21 is a neural network model that includes a first subset network 211, a second subset network 212, and a third subset network 213. The first subset network 211 is a first feature extraction layer that accepts input of a first compound structure and extracts features of the first compound structure. The second subset network 212 is a second feature extraction layer that accepts input of a second compound structure and extracts features of the second compound structure. The first subset network 211 and the second subset network 212 are independent of each other. The first subset network 211 and the second subset network 212 are generated such that the output from the similarity determination model 21 (i.e., similarity score) remains the same even if the input order of the first and second compound structures is changed. The third subset network 213 is a similarity representation layer that determines the similarity score based on the outputs from the first subset network 211 and the second subset network 212, respectively.

[0029] Known metrics used to obtain the similarity of compounds include the Tanimoto coefficient and Fraggle. However, since these metrics are predetermined, they may not take into account the substructure of compounds that users consider important as problem-specific factors. For example, if the same similarity metric is used for a system where differences in terminal functional groups are very important and for a system where the skeletal structure is important and terminal functional groups are not so important, the results may contradict the user's perception. In contrast, the similarity calculated by the similarity judgment model 21 (for example, the similarity judgment model 21 which is a neural network model) is generated by machine learning, making it possible to obtain results regarding similarity that match the user's perception.

[0030] (Selection of proposed compounds) The operation of the compound suggestion system 10 will be explained with reference to Figure 3. Figure 3 is a flowchart showing an example of the operation of the compound suggestion system 10 as processing flow S1.

[0031] In step S11, the initial data acquisition unit 12 acquires the number of proposals q and the dataset. In one example, the user terminal 40 generates initial data including the number of proposals q and the dataset in response to a predetermined user operation and transmits the initial data to the compound proposal system 10. The initial data acquisition unit 12 receives the initial data from the user terminal 40. The number of proposals q is an integer of 2 or more. The dataset contains multiple data records that show the compound structure and properties of the sample compound. Typically, the number of data records in the dataset may depend on the number of compounds experimented on, and is therefore less than (e.g., significantly less than) the number of candidate compounds indicated by the compound data. For example, the number of data records in the dataset may be 10 or several tens. At least one of the multiple sample compounds may also be a candidate compound indicated by the compound data.

[0032] In step S12, the second learning unit 13 performs regression analysis on the acquired dataset to generate a regression model 22. As described above, the regression model 22 includes a model function, uncertainty of the predicted values, and an acquisition function. The second learning unit 13 may generate a regression model based on a Gaussian process regression, a random forest, or a neural network as the regression model 22 for calculating the uncertainty of the predicted values.

[0033] In step S13, the penalty setting unit 14a initializes the penalty in the acquisition function for each of the multiple candidate compounds indicated by the compound data. A penalty is set for each candidate compound. For example, as an initialization process, the penalty setting unit 14a sets an initial value for the penalty so that the penalty is not substantially applied to each candidate compound.

[0034] In step S14, the compound selection unit 14b calculates the value of the penalty-based acquisition function for each of the multiple candidate compounds. The acquisition function is expressed as a function f(μ,δ,Φ) defined by the mean μ represented by the model function, the variance δ which indicates the uncertainty of the predicted value, and the penalty Φ.

[0035] For example, the function f(μ,δ,Φ) may be the acquisition function a(x) expressed by equation (1). The acquisition function a(x) is obtained by multiplying the acquisition function α0(x), which is defined by the mean μ and variance δ without considering the penalty Φ, by the penalty Φ, which is expressed by the product Π of the function φ(d).

number

[0036] The function φ(d) may also be the radial basis function (RBF) shown in equation (2).

number

[0037] When using Gaussian process regression as the regression model 22, the function φ(d) may be the local penalizer described in Reference 1 below. This local penalizer is a penalty function based on the Lipschitz constant. (Reference 1) Javier Gonzalez, Zhenwen Dai, Neil D. Lawrence, “Batch Bayesian Optimization via Local Penalization”, arXiv:1505.08052v2, 1 Jun 2015.

[0038] In step S15, the compound selection unit 14b selects one candidate compound as the proposed compound, which has the maximum value of the calculated acquisition function.

[0039] In step S16, the search unit 14 determines whether or not it has selected q proposed compounds. If the number of proposed compounds is less than q (NO in step S16), the process proceeds to step S17.

[0040] In step S17, the distance calculation unit 14c uses the intermediate layer of the similarity determination model 21 to calculate the distance between the compound structure of the last selected proposed compound and the compound structures of each of the multiple candidate compounds shown by the compound data. The "last selected proposed compound" refers to the proposed compound selected in the previous step S15.

[0041] When the similarity determination model 21 shown in Figure 2 is used, the distance calculation unit 14c uses the portion including the first subset network 211 and the second subset network 212, but not the third subset network 213, as an intermediate layer of the similarity determination model 21. In one example, the distance calculation unit 14c inputs the compound structure of the last selected proposed compound into the first subset network 211 and the compound structure of one compound into the second subset network 212. The distance calculation unit 14c calculates the distance based on the output from the first subset network 211 and the output from the second subset network 212. For example, the distance calculation unit 14c may calculate the Euclidean distance between the two outputs. In one example, the distance calculation unit 14c obtains first output data representing the features of the proposed compound from the first subset network 211 and second output data representing the features of candidate compounds from the second subset network 212. Then, the distance calculation unit 14c calculates the distance between the first output data and the second output data, i.e., the distance between the features. As mentioned above, the distance is, for example, the Euclidean distance. The distance calculation unit 14c calculates the distance for each of the multiple candidate compounds. The multiple candidate compounds shown by the compound data include the last selected proposed compound. Therefore, the distance calculation unit 14c also performs the process of inputting the feature quantities of the proposed compound into both the first subset network 211 and the second subset network 212 and calculating the distance, and this distance becomes 0.

[0042] In step S18, the penalty setting unit 14a updates the penalty for each of the multiple candidate compounds based on the distance to the candidate compound. The penalty setting unit 14a updates the penalty for each candidate compound so that the penalty increases as the distance decreases. For each of the one or more selected proposed compounds, the penalty setting unit 14a updates the penalty for the proposed compound so that the value of the acquisition function for the proposed compound is not maximized. For example, the penalty setting unit 14a updates the penalty for each proposed compound to a value corresponding to infinity. By setting the penalty for proposed compounds in this way, it is possible to avoid a situation where one candidate compound is selected as a proposed compound more than once. For example, the penalty setting unit 14a performs these penalty updates by recalculating the total product Π shown in equation (1) above.

[0043] After step S18, the process returns to step S14. In the repeated step S14, the compound selection unit 14b calculates the value of the acquisition function based on the updated penalty for each of the multiple candidate compounds. In the repeated step S15, the compound selection unit 14b selects the candidate compound with the highest calculated acquisition function value as the proposed compound.

[0044] If q proposed compounds are selected (YES in step S16), the process proceeds to step S19. In step S19, the recommendation unit 15 outputs q proposed compounds. The recommendation unit 15 generates recommendation information indicating the q proposed compounds and transmits this recommendation information to the user terminal 40. For example, the recommendation unit 15 may generate and transmit recommendation information indicating the compound structure of each proposed compound. The user terminal 40 receives and displays the recommendation information. The user can refer to the recommendation information and conduct experiments to obtain the properties of each of the q proposed compounds.

[0045] As shown in processing flow S1, the search unit 14 performs a search process using the regression model 22 for each compound structure of multiple candidate compounds stored in the compound database 32, and selects two or more candidate compounds from the multiple candidate compounds as proposed compounds. The search process includes repeating the compound selection step (step S15), the distance calculation step (step S17), and the penalty update step (step S18). The compound selection step is a step in which one candidate compound is selected as a proposed compound from multiple candidate compounds based on the acquisition function. The distance calculation step is a step in which the compound structure of the selected proposed compound is calculated to be the distance between each of the compound structures of the multiple candidate compounds, using the intermediate layers (first partial network 211 and second partial network 212) of the similarity determination model 21. The penalty update step is a step in which, for each of the multiple candidate compounds, the penalty corresponding to the candidate compound is updated based on the distance corresponding to the candidate compound.

[0046] Figure 4 illustrates an example of the search process in processing flow S1. In this example, assuming that the compound data consists of five candidate compounds C1, C2, C3, C4, and C5, and the number of proposals q is 3, the iterative process from steps S14 to S18 is explained. Figure 4 shows the change in the value of the acquisition function during the iteration. Figure 4 schematically represents regression model 200 as an example of regression model 22. The horizontal axis of the graph represents the search space (i.e., the range that the explanatory variable can take) composed of the compound structures of the multiple candidate compounds. The vertical axis of the graph represents the characteristics of the candidate compounds (i.e., the dependent variable). Regression model 200 includes the model function 201 (mean μ), the range 202 (variance δ) which shows the uncertainty of the predicted values, and the acquisition function 203.

[0047] In the first loop, the compound selection unit 14b calculates the value of the acquisition function for each of the candidate compounds C1 to C5 (step S14). In this example, it is assumed that the acquisition function value for candidate compound C2 is the maximum. In response to this result, the compound selection unit 14b selects candidate compound C2 as the proposed compound (step S15). The distance calculation unit 14c uses the intermediate layer of the similarity determination model 21 to calculate the distance between candidate compound (proposed compound) C2 and each of the candidate compounds C1 to C5 (step S17). The penalty setting unit 14a updates the penalty for each of the candidate compounds C1 to C5 (step S18). For example, the penalty setting unit 14a updates the penalty for candidate compound (proposed compound) C2 to a value corresponding to infinity.

[0048] In the second loop, the compound selection unit 14b calculates the value of the acquisition function for each of the candidate compounds C1 to C5 (step S14). In this example, it is assumed that the acquisition function value for candidate compound C4 is the maximum. In response to this result, the compound selection unit 14b selects candidate compound C4 as the proposed compound (step S15). The distance calculation unit 14c uses the intermediate layer of the similarity determination model 21 to calculate the distance between candidate compound (proposed compound) C4 and each of the candidate compounds C1 to C5 (step S17). The penalty setting unit 14a updates the penalty for each of the candidate compounds C1 to C5 (step S18). For example, the penalty setting unit 14a updates the penalty for candidate compound (proposed compound) C4 to a value corresponding to infinity. The penalty setting unit 14a maintains the penalty for candidate compound (proposed compound) C2.

[0049] In the third loop, the compound selection unit 14b calculates the value of the acquisition function for each of the candidate compounds C1 to C5 (step S14). In this example, it is assumed that the acquisition function value for candidate compound C1 is the largest. In response to this result, the compound selection unit 14b selects candidate compound C1 as the proposed compound (step S15). Since three proposed compounds have been selected, the search unit 14 terminates the search process (YES in step S16).

[0050] As shown in Figure 4, the mean μ represented by the model function 201 and the uncertainty (variance σ) of the predicted value indicated by the range 202 are determined when the second learning unit 13 generates the regression model 22 and are not changed in the subsequent search process. The value of the acquisition function for each candidate compound may change as the penalty is updated.

[0051] The processing flow S1 can be executed repeatedly. For example, the user conducts an experiment to obtain the properties of each of the q suggested compounds indicated in the recommendation information. As a result of this experiment, the user obtains the properties of one or more suggested compounds. The user adds one or more data records showing the compound structure and properties of one or more suggested compounds to the dataset. Thus, one or more suggested compounds are processed as one or more sample compounds. The compound suggestion system 10 obtains initial data including the number of suggestions q and the updated dataset (step S11), and then performs the subsequent processing (steps S12 to S19). As a result, the user obtains new q suggested compounds and can conduct further experiments (explorations). The number of suggestions q may or may not change each time the processing flow S1 is repeated.

[0052] [Differentiation] The technology relating to this disclosure has been described in detail above based on various examples. However, this disclosure is not limited to the examples given above. The technology relating to this disclosure can be modified in various ways without departing from its essence.

[0053] In the example above, the compound proposal system 10 generates the similarity determination model 21, but the compound proposal system may use a similarity determination model generated by another computer. That is, the similarity determination model may be ported from another computer system to the compound proposal system. Therefore, the compound proposal system does not need to have a function corresponding to the first learning unit 11.

[0054] In the example above, the compound suggestion system 10 acts as a server in a client-server system. In another example, the functions of the compound suggestion system 10, the learning database 31, and the compound database 32 may be implemented on a standalone computer. Alternatively, the compound suggestion system may be implemented on a user terminal that can access the learning database 31 and the compound database 32 via a communication network.

[0055] The processing steps for a method executed by at least one processor are not limited to the examples above. For example, some of the steps described above may be omitted, or each step may be performed in a different order. Also, any two or more of the steps described above may be combined, or some of the steps may be modified or deleted. Alternatively, other steps may be performed in addition to each of the steps described above.

[0056] In comparing the relative magnitudes of two numerical values ​​in this disclosure, either of the two criteria, "greater than or equal to" and "greater than," may be used, or either of the two criteria, "less than or equal to" and "less than," may be used.

[0057] In this disclosure, the expression "at least one processor executes a first process, a second process, ... and the nth process," or a corresponding expression, refers to a concept that includes cases where the entity executing the n processes from the first process to the nth process, i.e., the processor, changes along the way. In other words, this expression refers to a concept that includes both cases where all n processes are executed by the same processor and cases where the processor changes at an arbitrary rate for the n processes.

[0058] [Note] As can be seen from the various examples above, this disclosure includes the following aspects: (Note 1) Equipped with at least one processor, The at least one processor, A dataset containing multiple data records showing the compound structure and properties of the sample compound was obtained. A regression analysis is performed on the dataset to generate a regression model that includes a model function showing the relationship between the compound structure and the properties, the uncertainty of the predicted values ​​by the model function, and an acquisition function based on the model function, the uncertainty, and the penalty corresponding to the candidate compound. The search process using the regression model is performed on the compound structure of each of the multiple candidate compounds stored in the database, and two or more candidate compounds are selected from the multiple candidate compounds as proposed compounds. The aforementioned search process, The steps include selecting one candidate compound from the plurality of candidate compounds as the proposed compound based on the acquisition function, A step of calculating the distance between the compound structure of the selected proposed compound and the compound structures of each of the plurality of candidate compounds, using an intermediate layer of a similarity determination model that accepts input of a first compound structure and a second compound structure and determines the similarity between the first compound structure and the second compound structure; For each of the plurality of candidate compounds, the step of updating the penalty corresponding to the candidate compound based on the distance corresponding to the candidate compound, Includes repetition of Compound suggestion system. (Note 2) In the selection step of the search process, the candidate compound with the maximum value of the acquisition function is selected as the proposed compound. In the updating step of the search process, the penalty for the selected proposed compound is updated so that the value of the acquisition function of the selected proposed compound is not maximized. The compound proposal system described in Appendix 1. (Note 3) The similarity determination model is a neural network model that includes a first subset network that accepts the input of the first compound structure, a second subset network that accepts the input of the second compound structure independently of the first subset network, and a third subset network that determines the similarity based on the outputs from the first subset network and the second subset network, respectively. The intermediate layer includes the first and second subset networks, but does not include the third subset network. The at least one processor, for each of the plurality of candidate compounds, The compound structure of the candidate compound is input into one of the first and second partial networks. The compound structure of the selected proposed compound is input to the other of the first and second partial networks. The distance is calculated based on the output from the first subset network and the output from the second subset network. The compound proposal system described in Appendix 1 or 2. (Note 4) The at least one processor generates the similarity determination model by performing machine learning using training data which includes a plurality of data records indicating the correspondence between pairs of compound structures and the similarity scores. A compound proposal system described in any one of the appendices 1 to 3. (Note 5) A compound proposal method performed by a compound proposal system comprising at least one processor, The steps include obtaining a dataset containing multiple data records that show the compound structure and properties of a sample compound, A step of performing regression analysis on the dataset to generate a regression model that includes a model function showing the relationship between the compound structure and the properties, uncertainty of the predicted values ​​by the model function, and an acquisition function based on the model function, the uncertainty, and a penalty corresponding to the candidate compound. The steps include performing the search process using the regression model described above on the compound structure of each of the multiple candidate compounds stored in the database, and selecting two or more candidate compounds from the multiple candidate compounds as proposed compounds, Includes, The aforementioned search process, The steps include selecting one candidate compound from the plurality of candidate compounds as the proposed compound based on the acquisition function, A step of calculating the distance between the compound structure of the selected proposed compound and the compound structures of each of the plurality of candidate compounds, using an intermediate layer of a similarity determination model that accepts input of a first compound structure and a second compound structure and determines the similarity between the first compound structure and the second compound structure; For each of the plurality of candidate compounds, the step of updating the penalty corresponding to the candidate compound based on the distance corresponding to the candidate compound, Includes repetition of Compound proposal method. (Note 6) The steps include obtaining a dataset containing multiple data records that show the compound structure and properties of a sample compound, A step of performing regression analysis on the dataset to generate a regression model that includes a model function showing the relationship between the compound structure and the properties, uncertainty of the predicted values ​​by the model function, and an acquisition function based on the model function, the uncertainty, and a penalty corresponding to the candidate compound. The steps include performing the search process using the regression model described above on the compound structure of each of the multiple candidate compounds stored in the database, and selecting two or more candidate compounds from the multiple candidate compounds as proposed compounds, Have the computer run it, The aforementioned search process, The steps include selecting one candidate compound from the plurality of candidate compounds as the proposed compound based on the acquisition function, A step of calculating the distance between the compound structure of the selected proposed compound and the compound structures of each of the plurality of candidate compounds, using an intermediate layer of a similarity determination model that accepts input of a first compound structure and a second compound structure and determines the similarity between the first compound structure and the second compound structure; For each of the plurality of candidate compounds, the step of updating the penalty corresponding to the candidate compound based on the distance corresponding to the candidate compound, Includes repetition of Compound proposal program.

[0059] According to appendices 1, 5, and 6, a regression model is generated that shows the relationship between the compound structure and properties of the compounds. Through a search process using this regression model, two or more candidate compounds are selected as proposed compounds from among multiple candidate compounds. The regression model includes a model function that shows the relationship between the compound structure and properties, an acquisition function based on the uncertainty of the predicted values ​​by the model function, and penalties corresponding to the candidate compounds. In the search process, each time a proposed compound is selected, the distance between the proposed compound and each candidate compound is calculated using an intermediate layer of a similarity determination model that determines the similarity between compound structures, and the penalty for each candidate compound is updated based on these distances. Therefore, when selecting the next proposed compound, the acquisition function based on the updated penalties is used. By reflecting the distance between compound structures obtained based on the intermediate layer of the similarity determination model in the acquisition function that directly contributes to the selection of proposed compounds, it is possible to propose two or more candidate compounds from among multiple candidate compounds in a manner that also takes into account the user's perception.

[0060] According to Appendix 2, the penalty for the selected proposed compound is updated so that the value of the acquisition function of the selected proposed compound is not maximized, thus ensuring that a different candidate compound is selected as the proposed compound. Therefore, the situation in which one candidate compound is selected more than once is avoided, and two or more proposed compounds can be reliably selected in a short computation time.

[0061] According to Appendix 3, the distance between the compound structure of the selected proposed compound and the compound structures of each candidate compound is calculated based on the outputs from the first and second subset networks, which are intermediate layers of the similarity determination model and are independent of each other. This mechanism using a neural network model allows for a more accurate calculation of the distance between compound structures, enabling the proposal of two or more candidate compounds in a manner that better considers the user's perception.

[0062] According to Appendix 4, the similarity determination model is generated by the compound suggestion system. Therefore, the process of suggesting two or more compounds from multiple compounds, taking into account the user's perception, can be completed without using any other computer system. [Explanation of Symbols]

[0063] 10... Compound suggestion system, 11... First learning unit, 12... Initial data acquisition unit, 13... Second learning unit, 14... Search unit, 14a... Penalty setting unit, 14b... Compound selection unit, 14c... Distance calculation unit, 15... Recommendation unit, 21... Similarity judgment model, 22... Regression model, 31... Learning database, 32... Compound database, 40... User terminal, 101... Processor, 200... Regression model, 211... First subset network, 212... Second subset network, 213... Third subset network.

Claims

1. Equipped with at least one processor, The at least one processor, A dataset containing multiple data records showing the compound structure and properties of the sample compound was obtained. A regression analysis is performed on the dataset to generate a regression model that includes a model function showing the relationship between the compound structure and the properties, the uncertainty of the predicted values ​​by the model function, and an acquisition function based on the model function, the uncertainty, and the penalty corresponding to the candidate compound. The search process using the regression model is performed on the compound structure of each of the multiple candidate compounds stored in the database, and two or more candidate compounds are selected from the multiple candidate compounds as proposed compounds. The aforementioned search process, The steps include selecting one candidate compound from the plurality of candidate compounds as the proposed compound based on the acquisition function, A step of calculating the distance between the compound structure of the selected proposed compound and the compound structures of each of the plurality of candidate compounds, using an intermediate layer of a similarity determination model that accepts input of a first compound structure and a second compound structure and determines the similarity between the first compound structure and the second compound structure; For each of the plurality of candidate compounds, the step of updating the penalty corresponding to the candidate compound based on the distance corresponding to the candidate compound, Includes repetition of Compound suggestion system.

2. In the selection step of the search process, the candidate compound with the maximum value of the acquisition function is selected as the proposed compound. In the updating step of the search process, the penalty for the selected proposed compound is updated so that the value of the acquisition function of the selected proposed compound is not maximized. The compound proposal system according to claim 1.

3. The similarity determination model is a neural network model that includes a first subset network that accepts input of the first compound structure, a second subset network that accepts input of the second compound structure independently of the first subset network, and a third subset network that determines the similarity based on the outputs from the first subset network and the second subset network, respectively. The intermediate layer includes the first and second subset networks, but does not include the third subset network. The at least one processor, for each of the plurality of candidate compounds, The compound structure of the candidate compound is input into one of the first and second partial networks. The compound structure of the selected proposed compound is input into the other of the first and second partial networks. The distance is calculated based on the output from the first subset network and the output from the second subset network. The compound proposal system according to claim 1 or 2.

4. The at least one processor generates the similarity determination model by performing machine learning using training data which includes a plurality of data records indicating the correspondence between pairs of compound structures and the similarity scores. The compound proposal system according to claim 1 or 2.

5. A compound proposal method performed by a compound proposal system comprising at least one processor, The steps include obtaining a dataset containing multiple data records that show the compound structure and properties of a sample compound, A step of performing regression analysis on the dataset to generate a regression model that includes a model function showing the relationship between the compound structure and the properties, uncertainty of the predicted values ​​by the model function, and an acquisition function based on the model function, the uncertainty, and a penalty corresponding to the candidate compound. The steps include performing the search process using the regression model described above on the compound structure of each of the multiple candidate compounds stored in the database, and selecting two or more candidate compounds from the multiple candidate compounds as proposed compounds, Includes, The aforementioned search process, The steps include selecting one candidate compound from the plurality of candidate compounds as the proposed compound based on the acquisition function, A step of calculating the distance between the compound structure of the selected proposed compound and the compound structures of each of the plurality of candidate compounds, using an intermediate layer of a similarity determination model that accepts input of a first compound structure and a second compound structure and determines the similarity between the first compound structure and the second compound structure; For each of the plurality of candidate compounds, the step of updating the penalty corresponding to the candidate compound based on the distance corresponding to the candidate compound, Includes repetition of Compound proposal method.

6. The steps include obtaining a dataset containing multiple data records that show the compound structure and properties of a sample compound, A step of performing regression analysis on the dataset to generate a regression model that includes a model function showing the relationship between the compound structure and the properties, uncertainty of the predicted values ​​by the model function, and an acquisition function based on the model function, the uncertainty, and a penalty corresponding to the candidate compound. The steps include performing the search process using the regression model described above on the compound structure of each of the multiple candidate compounds stored in the database, and selecting two or more candidate compounds from the multiple candidate compounds as proposed compounds, Have the computer run it, The aforementioned search process, The steps include selecting one candidate compound from the plurality of candidate compounds as the proposed compound based on the acquisition function, A step of calculating the distance between the compound structure of the selected proposed compound and the compound structures of each of the plurality of candidate compounds, using an intermediate layer of a similarity determination model that accepts input of a first compound structure and a second compound structure and determines the similarity between the first compound structure and the second compound structure; For each of the plurality of candidate compounds, the step of updating the penalty corresponding to the candidate compound based on the distance corresponding to the candidate compound, Includes repetition of Compound proposal program.

Images