Computer system and character string generation method

Abstract

This computer system receives, as an input, a character string in which at least one position is masked, holds information on a generation model that outputs a probability distribution for each character at the masked position, receives the input of a target character string, generates a plurality of verification character strings in which one position of the target character string is masked, inputs the plurality of verification character strings into the generation model, acquires the probability distribution for each of the plurality of verification character strings, determines a mutation position of the target character string on the basis of the probability distribution, and determines a character at the mutation position on the basis of the probability distribution acquired by inputting a masked character string in which the mutation position of the target character string is masked into the generation model, thereby generating a mutated character string.

Description

Computer system, character string generation method Incorporation by Reference

[0001] This application claims priority from Japanese Patent Application No. 2024-90956, filed on June 4, 2024, the contents of which are incorporated herein by reference.

[0002] The present invention relates to a method for generating an array having desired function values.

[0003] In protein design, there is a technique described in Patent Document 1 for discovering amino acid sequences with desired functions. Patent Document 1 states that "part of a mutant library prepared by randomly introducing mutations is evaluated for two or more properties."

[0004] International Publication No. 2023 / 047557 Special Publication No. 2022-543234

[0005] In Patent Document 1, mutants created by randomly introducing mutations are evaluated, so if the search space is huge and there are many mutation candidates, there is a high probability that mutants with desired properties will not be included. An object of the present invention is to efficiently search for highly functional amino acid sequences.

[0006] A representative example of the invention disclosed in the present application is as follows: That is, a computer system that converts at least one character of a string consisting of multiple characters to generate a new string includes a processor and a storage device connected to the processor, the storage device stores information about a generative model that accepts as input a string with at least one masked portion and outputs a probability distribution of each character in the masked portion, the processor accepts as input a target string, generates multiple verification strings in which one portion of the target string is masked, inputs the multiple verification strings to the generative model to obtain the probability distribution for each of the multiple verification strings, determines a mutation portion of the target string based on the probability distribution, and generates a mutant string by inputting a masked string in which the mutation portion of the target string is masked to the generative model and determining the character at the mutation portion based on the probability distribution obtained.

[0007] According to the present invention, it is possible to efficiently search for mutants that improve function while maintaining a partial sequence pattern. Objects, configurations, and effects other than those described above will become clear from the following description of the examples.

[0008] FIG. 1 is a diagram illustrating an example of a hardware configuration of a sequence generation system according to a first embodiment. FIG. 2 is a diagram illustrating an example of functional blocks of the sequence generation system according to the first embodiment. FIG. 3 is a flowchart illustrating a flow of processing executed by the sequence generation system according to the first embodiment. FIG. 4 is a diagram illustrating a generation process of a generation model configuration and a probability distribution according to the first embodiment. FIG. 5 is a flowchart illustrating an example of a learning process of a function prediction model executed by the sequence generation system according to the first embodiment. FIG. 6 is a flowchart illustrating an example of a learning process of a generation model executed by the sequence generation system according to the first embodiment. FIG. 7 is a flowchart illustrating an example of a mutation site determination process executed by the sequence generation system according to the first embodiment. FIG. 8 is a flowchart illustrating an example of a mutant sequence generation process executed by the sequence generation system according to the first embodiment. FIG. 9 is a flowchart illustrating an example of a mutant sequence evaluation process executed by the sequence generation system according to the first embodiment. FIG. 10 is a flowchart illustrating a flow of processing executed by the sequence generation system according to the second embodiment.

[0009] In the field of protein design, the technology described in Patent Document 2 is also known. Patent Document 2 describes a "system comprising a supervised model for predicting function and a decoder network," and "the system is trained to provide probabilistic biopolymer sequences given the embedding of biopolymer sequences in a function space." Patent Document 2 does not anticipate generating mutant sequences while maintaining specific portions of the base sequence for the biopolymer sequences provided.

[0010] In the following examples, a system for efficiently searching for highly functional amino acids while maintaining amino acids at specific positions in a sequence will be described. The examples will be described below with reference to the accompanying drawings.

[0011] Below, an example will be described that assumes pharmaceutical development.

[0012] First, a configuration example of a sequence generation system 100 will be described with reference to Figures 1 and 2. Figure 1 is a diagram illustrating an example of a hardware configuration of the sequence generation system 100 according to the first embodiment. Also, Figure 2 is a diagram illustrating an example of functional blocks of the sequence generation system 100 according to the first embodiment.

[0013] The sequence generation system 100 of the first embodiment is a computer system that generates an amino acid sequence representing a protein, and is realized by a general-purpose computer or a server device having the components described below. The amino acid sequence is expressed as a sequence of characters corresponding to the type of amino acid. Hereinafter, this will be referred to as a sequence.

[0014] The sequence generation system 100 is connected to other computers, server devices, etc., from which input data 201 is obtained, via a communication network such as the Internet or a dedicated line, so that data can be exchanged between them.

[0015] The sequence generation system 100 and other devices are connected to the communication network by wire via well-known communication equipment, but may also be connected wirelessly.

[0016] In this embodiment, the sequence generation system 100 has been described as being composed of one device, but it may be composed of multiple devices.

[0017] Next, the hardware configuration of the sequence generation system 100 will be described.

[0018] 1, the sequence generation system 100 is realized by a single general-purpose computer that includes one or more processors 101, one or more storage devices 102, a plurality of interface devices 103, 104, and 105, and a wired or wireless communication line 106 connecting them. Note that the sequence generation system 100 may also be realized using a plurality of general-purpose computers.

[0019] Next, we will explain the functions of the array generation system 100. Note that each block explained below does not represent a hardware configuration, but represents a functional block.

[0020] The sequence generation system 100 includes a sequence generation unit 202, a sequence evaluation unit 203, and a storage unit 204. The sequence generation system 100 also includes a user interface unit (not shown) that is composed of a communication unit, an input unit, and an output unit.

[0021] An input unit (not shown) receives, as input data 201 , a sequence 211 , a mutation introduction rate 212 , constraint information 213 , and setting information 214 .

[0022] The mutation rate 212 is an example of a condition related to the degree of mutation in a sequence, and may be information indicating the degree of mutation in a sequence, such as the number of mutation sites in the sequence. The mutation rate 212 may also be an index such as the edit distance of the sequence.

[0023] The constraint information 213 is information relating to constraints on sequence mutations. The constraint information 213 includes, for example, parts of the sequence that should not be changed, the distance relationship between amino acids in the sequence, etc. However, the constraint information 213 is not limited to the above.

[0024] An output unit (not shown) outputs output data 205. The output data 205 includes a mutant sequence 251 and execution history data 252. The execution history data 252 is data indicating the results of various processes performed by the sequence generation system 100, and includes, for example, learning results.

[0025] The sequence generation unit 202 executes various processes for generating a mutant sequence 251. Specifically, the sequence generation unit 202 executes supervised learning processing of the function prediction model 222 ( FIGS. 3 and 5 ), reinforcement learning processing of the generative model 227 ( FIGS. 3 and 6 ), mutation site determination processing ( FIGS. 3 and 7 ), and mutant sequence generation processing ( FIGS. 3 and 8 ). The sequence generation unit 202 also functions as an interface for the storage unit 204, the sequence evaluation unit 203, and the communication unit.

[0026] As shown in FIG. 2, the sequence generation unit 202 includes a function prediction model improvement unit 220, a generation model improvement unit 225, and a mutation site determination unit 230.

[0027] The function prediction model improving unit 220 includes a function predicting unit 221 and a function prediction model updating unit 224. The function prediction model improving unit 220 executes a learning process for the function prediction model 222.

[0028] The function prediction unit 221 includes the function blocks of a function prediction model 222 and a predicted value conversion unit 223. The function prediction unit 221 receives an array as input and outputs a predicted value of the function value of the array.

[0029] The function prediction model 222 is a machine learning model such as a neural network, which receives a sequence as input and outputs a probability distribution representing the level of function of each amino acid that constitutes the sequence.

[0030] The predicted value conversion unit 223 converts the probability distribution output by the function prediction model 222 into a predicted function value of the array.

[0031] The function prediction model update unit 224 updates the function prediction model 222 using the array with function values ​​and the predicted function values ​​output by the function prediction unit 221. Here, the array with function values ​​is an array for learning.

[0032] The generative model improvement unit 225 includes a mutant sequence generation unit 226 and a generative model update unit 229. The generative model improvement unit 225 executes a learning process for the generative model 227.

[0033] The mutant sequence generation unit 226 includes a generation model 227 and a sequence conversion unit 228. The mutant sequence generation unit 226 receives as input a mask sequence in which the mutation site determined by the mutation site determination unit 230 has been replaced with a mask token, which is a special symbol, and outputs a mutant sequence 206 in which the mask token has been replaced with an arbitrary amino acid. Note that, in this embodiment, a mutant sequence having the same length as the input sequence is generated, but a mutant sequence having a length different from that of the input sequence may also be generated.

[0034] The generative model 227 is a machine learning model such as a neural network, which receives as input a masked sequence in which part of a sequence with function values ​​has been replaced with a masked token, and outputs a probability distribution of each amino acid in the masked portion.

[0035] The sequence conversion unit 228 outputs the mutant sequence 206 based on the probability distribution output by the generative model 227 .

[0036] The generative model update unit 229 updates the generative model 227 using the array with the function values ​​and the predicted function values ​​output by the function prediction unit 221 .

[0037] The mutation site determination unit 230 generates a candidate mutant sequence in which one position of the input sequence 211 is replaced with a mask token, sends it to the generation model 227, and determines the mutation site of the sequence 211 based on the probability distribution of amino acids at the position replaced with the mask token output by the generation model 227.

[0038] The sequence evaluation unit 203 evaluates the functional values ​​of the mutant sequence 206 output by the sequence generation unit 202, and outputs the mutant sequence 206 with the functional values ​​as an evaluation result 207 (FIGS. 3 and 9). The evaluation result 207 is stored in the sequence store 241 of the storage unit 204.

[0039] Next, a description will be given of the processing of the sequence generation system 100. Fig. 3 is a flowchart showing the flow of processing executed by the sequence generation system 100 of the first embodiment.

[0040] In step S301, the function prediction model improving unit 220 executes a learning process for the function prediction model 222.

[0041] Specifically, the function prediction model improvement unit 220 updates the function prediction model 222 so as to minimize the error between the predicted function value obtained by inputting an array with function values ​​to the function prediction unit 221 and the function value of the array with function values. This enables the function prediction model 222 to predict function values ​​for an array.

[0042] The function prediction model 222 may be updated, for example, by learning so that the function prediction model 222 outputs information on the superiority or inferiority of sequences.

[0043] In step S302, the generative model improvement unit 225 executes a learning process for the generative model 227.

[0044] In this embodiment, the generative model 227 is trained using a reinforcement learning technique. Specifically, the generative model improvement unit 225 obtains a mutant sequence by inputting a sequence with a function value to the mutant sequence generation unit 226. The generative model 227 outputs a probability distribution that takes into account the degree of association between amino acids in the sequence. The generative model improvement unit 225 inputs the mutant sequence to the function prediction unit 221 and obtains a predicted function value of the mutant sequence. The generative model improvement unit 225 updates the generative model 227 so that the predicted function value is maximized.

[0045] By repeating the generation of mutant sequences and the prediction of function values, and updating the generative model 227 so as to improve the predicted function values, it becomes possible to efficiently search for highly functional sequences.

[0046] In step S303, the mutation site determining unit 230 receives the input of the sequence 211 and executes a mutation site determining process.

[0047] Specifically, the mutation site determination unit 230 identifies all positions in the sequence 211 where mutations can be introduced based on the constraint information 213, replaces one of the identified positions with a mask token, and generates a verification sequence. This process generates a number of verification sequences, each with one of the positions where mutations can be introduced masked, equal to the number of positions where mutations can be introduced. The mutation site determination unit 230 inputs the verification sequence into the generation model 227 and obtains a probability distribution of amino acids for the sequence position. The mutation site determination unit 230 calculates the maximum probability of the masked position based on the probability distribution and determines the mutation site based on the maximum probability. For example, the mutation site determination unit 230 determines the masked position in the verification sequence whose maximum probability is greater than a threshold as the mutation site. The number of mutations to be introduced is determined based on the mutation introduction rate 212. For example, the mutation site determination unit 230 determines the mutation site in descending order of maximum probability until the mutation rate reaches the mutation introduction rate 212. The mutation site determination unit 230 transmits the masked sequence in which the determined mutation sites have been replaced with mask tokens to the mutation sequence generation unit 226.

[0048] The above-described method for determining the mutation location is an example and is not limiting. For example, the mutation location may be determined in ascending order of minimum probability. Alternatively, some locations where mutations can be introduced may be selected randomly. Data may be transmitted and received between the mutation location determination unit 230 and the generative model 227 multiple times.

[0049] In step S304, the variant sequence generating unit 226 executes a variant sequence generating process.

[0050] Specifically, the mutant sequence generation unit 226 generates a mutant sequence by inputting the mask sequence into the generation model 227. By using the generation model 227, a mutant sequence is generated taking into consideration the degree of association between the mutation site and the site where no mutation is introduced.

[0051] Here, the generative model 227 is trained to maximize the predicted function value, and therefore mutations (amino acids) are introduced at mutation sites to create highly functional sequences. Furthermore, because the mutation sites are determined taking into account the constraint information 213, the mutated sequences satisfy the constraints. This makes it possible to efficiently search for proteins by maintaining a specific sequence pattern and introducing mutations at sites where there is a high possibility of improving function.

[0052] In step S305, the sequence evaluation unit 203 executes an evaluation process for the mutant sequence 206. The sequence evaluation unit 203 stores the evaluation result 207 obtained by the evaluation process in the storage unit 204. The sequence evaluation method uses a neural network to evaluate the functional value of the mutant sequence. The evaluation method may also be the result of an experiment conducted on a prototype mutant sequence.

[0053] In step S306, the sequence generation system 100 counts the number of times the sequence is evaluated, and then determines whether the number is greater than a predetermined value. If the number of times the sequence is evaluated is equal to or less than the predetermined value, the sequence generation system 100 returns to step S301 and executes the same processing. If the number of times the sequence is evaluated is greater than the predetermined value, the sequence generation system 100 terminates the processing. The sequence generation system 100 transmits the final mutant sequence 251 and the execution history data 252 to the output unit.

[0054] Next, a description will be given of the generative model 227. Fig. 4 is a diagram showing the configuration of the generative model 227 and the process of generating a probability distribution according to the first embodiment.

[0055] The generative model 227 is composed of a relevance calculation unit 401 and a probability distribution conversion unit 402. The generative model 227 receives a mask sequence as input. The mask sequence is a sequence in which mutation locations are replaced with mask tokens.

[0056] The relevance calculation unit 401 calculates an embedded representation of the mask sequence, and calculates three vectors, a query, a key, and a value, using information obtained by adding positional information of each amino acid in the sequence to the embedded representation of the mask sequence.The relevance calculation unit 401 calculates an attention score using the query, the key, and the number of dimensions of the query, and calculates feature amounts from the attention score and the value.This makes it possible to determine the relevance between the mutation position and each amino acid in the sequence.

[0057] The probability distribution conversion unit 402 converts the calculated feature amount into a probability distribution of amino acids at the mutation site, and outputs the probability distribution.

[0058] In this way, by determining mutation sites in a sequence using the output of the generative model 227 that takes into account the relatedness between amino acids in the sequence, it is possible to narrow down the sequence patterns to be searched for, thereby enabling efficient searching for highly functional sequences.

[0059] Next, each process executed by the sequence generation system 100 of the first embodiment will be described with reference to FIGS.

[0060] FIG. 5 is a flowchart illustrating an example of a learning process of the function prediction model 222 executed by the sequence generation system 100 according to the first embodiment.

[0061] In step S501, the function prediction model improving unit 220 acquires, via the input unit or the communication unit, the setting information 214. The setting information 214 includes information necessary for the learning process, such as hyperparameters.

[0062] In step S502 , the function prediction model improvement unit 220 acquires an array with function values ​​from the array store 241 , and acquires the function prediction model 222 from the function prediction model store 242 .

[0063] In step S503 , the function prediction model improving unit 220 performs supervised learning of the function prediction model 222 .

[0064] In step S504, the function prediction model improvement unit 220 stores the learned function prediction model 222 in the function prediction model store 242. Thereafter, the function prediction model improvement unit 220 ends the learning process of the function prediction model 222.

[0065] FIG. 6 is a flowchart illustrating an example of a learning process of the generation model 227 executed by the sequence generation system 100 according to the first embodiment.

[0066] In step S601, the generative model improvement unit 225 acquires, via the input unit or the communication unit, the setting information 214. The setting information 214 includes information necessary for the learning process, such as hyperparameters.

[0067] In step S602 , the generative model improvement unit 225 acquires an array with function values ​​from the array store 241 , acquires the function prediction model 222 from the function prediction model store 242 , and acquires the generative model 227 from the generative model store 243 .

[0068] In step S603, the generative model improving unit 225 determines parameters for the prediction value converting unit 223 based on the setting information 214, and sets a performance evaluation index. In this way, a performance evaluation index in reinforcement learning is set.

[0069] In step S604, the generative model improvement unit 225 performs reinforcement learning on the generative model 227 so as to maximize the performance evaluation index.

[0070] In step S605, the generative model improvement unit 225 stores the learned generative model 227 in the generative model store 243. Thereafter, the generative model improvement unit 225 ends the learning process of the generative model 227.

[0071] FIG. 7 is a flowchart illustrating an example of a mutation site determination process executed by the sequence generation system 100 according to the first embodiment.

[0072] In step S701, the mutation site determining unit 230 acquires the sequence 211, the mutation introduction rate 212, the constraint information 213, and the setting information 214 via the input unit or the communication unit.

[0073] In step S702, the mutation location determination unit 230 acquires the generative model 227 from the generative model store 243.

[0074] In step S703, the mutation site determining unit 230 generates a verification sequence based on the sequence and constraint information 213.

[0075] In step S704, the mutation site determining unit 230 inputs the verification sequences into the generative model 227, and obtains the probability distribution of amino acids at the mutation site for each verification sequence.

[0076] In step S705, the mutation site determining unit 230 determines a mutation site based on the probability distribution.

[0077] In step S706, the mutation location determination unit 230 generates a masked sequence by replacing the determined mutation location in the sequence 211 with a mask token. The mutation location determination unit 230 transmits the masked sequence to the mutation sequence generation unit 226. Thereafter, the mutation location determination unit 230 ends the mutation location determination process.

[0078] FIG. 8 is a flowchart illustrating an example of a mutant sequence generation process executed by the sequence generation system 100 according to the first embodiment.

[0079] In step S801, the mutation sequence generating unit 226 acquires a mask sequence from the mutation site determining unit 230.

[0080] In step S802, the mutation sequence generation unit 226 inputs the mask sequence to the generative model 227 and obtains the probability distribution of amino acids at the mutation site.

[0081] In step S803, the mutant sequence generating unit 226 generates the mutant sequence 206 by inputting the probability distribution of amino acids at the mutation site to the sequence converting unit 228.

[0082] In step S804, the variant sequence generating unit 226 transmits the generated variant sequence to the sequence evaluating unit 203. Thereafter, the variant sequence generating unit 226 ends the variant sequence generating process.

[0083] FIG. 9 is a flowchart illustrating an example of a mutation sequence evaluation process executed by the sequence generation system 100 according to the first embodiment.

[0084] In step S901, the sequence evaluation unit 203 acquires the mutant sequence 206 from the sequence generation unit 202.

[0085] In step S902, the sequence evaluation unit 203 acquires the functional value of the mutant sequence 206. For example, the functional value of the mutant sequence is acquired by synthesizing a protein of the mutant sequence using an experimental device and measuring the functional value of the synthesized protein. Note that the method is not limited to a method using an experimental device, and any process can be used as long as it is possible to acquire a pair of a mutant sequence and its functional value by synthesizing a mutant sequence and measuring its functional value. The functional value of the mutant sequence may also be acquired by simulation.

[0086] In step S903, the sequence evaluation unit 203 stores the mutant sequence 206 and the evaluation result 207 including the function value in the sequence store 241 of the storage unit 204. Thereafter, the sequence evaluation unit 203 ends the mutant sequence evaluation process.

[0087] According to the first embodiment, highly functional sequences can be efficiently searched for while satisfying sequence constraints.

[0088] In the second embodiment, the mutation location determination unit 230 includes a mutation location determination model and a mutation location determination model update unit, and updates the mutation location determination model according to the number of iterations.

[0089] The hardware configuration of the sequence generation system 100 of Example 2 is the same as that of Example 1. The functional blocks of the sequence generation system 100 of Example 2 are partially different from those of Example 1. Specifically, the mutation site determination unit 230 includes a mutation site determination model and a mutation site determination model update unit. The mutation site determination model is stored in a mutation site determination model store in the storage unit 204.

[0090] FIG. 10 is a flowchart showing the flow of processing executed by the sequence generation system 100 of the second embodiment.

[0091] In step S1001 , the function prediction model improving unit 220 executes a learning process for the function prediction model 222 .

[0092] In step S1002 , the generative model improvement unit 225 executes a learning process for the generative model 227 .

[0093] In step S1003, the mutation site determining unit 230 executes a mutation site determining process.

[0094] Specifically, the mutation site determination unit 230 reads out the mutation site determination model from the mutation site determination model store, and reads out the sequence with the function value from the sequence store 241. The mutation site determination unit 230 inputs the received sequence and the read out sequence with the function value into the mutation site determination model, and determines the mutation site.

[0095] In step S1004, the variant sequence generating unit 226 executes a variant sequence generating process.

[0096] In step S1005 , the sequence evaluation unit 203 executes an evaluation process for the mutant sequence 206 .

[0097] In step S1006 , the mutation location determination unit 230 calculates a performance evaluation index of the mutation location determination model based on the evaluation result 207 .

[0098] In step S1007, the mutation location determination model update unit executes a learning process for the mutation location determination model.

[0099] Specifically, the mutation site determination model update unit uses the received sequence, the sequence with the function value acquired from the sequence store 241, the mutation site determined in step S1103, and the performance evaluation index calculated in step S1107 to update the mutation site determination model using a reinforcement learning algorithm so as to improve the performance evaluation index. The mutation site determination model update unit stores the updated mutation site determination model in the mutation site determination model store of the storage unit 204.

[0100] In step S1008, the sequence generation system 100 counts the number of times the sequence is evaluated, and then determines whether the number is greater than a predetermined value.

[0101] In the third embodiment, the function prediction model 222 is utilized in the mutant sequence generation process.

[0102] Only steps that are changed from the first embodiment will be described with reference to FIG.

[0103] In step S803, the variant sequence generation unit 226 transmits the probability distribution to the sequence conversion unit 228 to generate a variant sequence 206. The variant sequence generation unit 226 transmits the generated variant sequence 206 to the function prediction unit 221 to obtain a predicted function value. The variant sequence generation unit 226 outputs a variant sequence 206 whose obtained predicted function value is close to the desired value. This process may be executed multiple times. When the process is completed, the process proceeds to step S804.

[0104] In Example 4, the method of Example 1 is applied to the design of compounds in material development.

[0105] This is realized by associating the amino acid sequence in Example 1 with the chemical formula sequence of the compound. By treating multiple chemical structures as one symbol, the introduction of a mask symbol and the generation of a mutant sequence 206 can be realized in the same way as in Example 1.

[0106] In the fifth embodiment, the technique of the first embodiment is applied to prompt design in image generation.

[0107] This is realized by associating the amino acid sequence in Example 1 with the sequence of a word in a natural language. The sequence evaluation unit 203 inputs the mutant sequence 206 into an image generator and obtains an evaluation value for the output image. The sequence evaluation unit 203 stores the evaluation result 207 in the sequence store 241 of the storage unit 204.

[0108] In the sixth embodiment, the technique of the first embodiment is applied to the generation of programming code.

[0109] This is realized by associating the amino acid sequence in Example 1 with a programming character string. The mutated sequence 206 output by the sequence generation unit 202 becomes programming code. The sequence evaluation unit 203 obtains an evaluation value for the mutated sequence 206 based on the execution result of the mutated sequence 206. The sequence evaluation unit 203 saves the evaluation result 207 in the sequence store 241 of the storage unit 204.

[0110] The present invention is not limited to the above-described embodiments, but includes various modifications. For example, the above-described embodiments are provided to explain the present invention in detail, and the present invention is not necessarily limited to those including all of the described configurations. Furthermore, some of the configurations of each embodiment can be added to, deleted from, or replaced with other configurations.

[0111] Furthermore, the above-described configurations, functions, processing units, processing means, etc. may be partially or entirely implemented in hardware, for example, by designing them as integrated circuits. The present invention can also be realized by software program code that implements the functions of the embodiments. In this case, a storage medium on which the program code is recorded is provided to a computer, and a processor included in the computer reads the program code stored in the storage medium. In this case, the program code itself read from the storage medium implements the functions of the above-described embodiments, and the program code itself and the storage medium on which it is stored constitute the present invention. Examples of storage media for providing such program code include flexible disks, CD-ROMs, DVD-ROMs, hard disks, solid-state drives (SSDs), optical disks, magneto-optical disks, CD-Rs, magnetic tape, non-volatile memory cards, and ROMs.

[0112] Furthermore, the program code that realizes the functions described in this embodiment can be implemented in a wide range of programming or scripting languages, such as assembler, C / C++, perl, Shell, PHP, Python, and Java.

[0113] Furthermore, the program code of the software that realizes the functions of the embodiments may be distributed via a network and stored in a storage means such as a computer's hard disk or memory, or in a storage medium such as a CD-RW or CD-R, and the processor of the computer may read and execute the program code stored in the storage means or the storage medium.

[0114] In the above-described embodiment, the control lines and information lines are shown as those considered necessary for the explanation, and not all control lines and information lines are necessarily shown in the product. All components may be interconnected.

Claims

1. A computer system that converts at least one character of a string consisting of multiple characters and generates a new string, comprising: a processor; and a storage device connected to the processor, wherein the storage device stores information about a generative model that accepts as input a string with at least one masked portion and outputs a probability distribution for each character in the masked portion, and wherein the processor accepts as input a target string; generates multiple verification strings in which one portion of the target string is masked; inputs the multiple verification strings to the generative model to obtain the probability distribution for each of the multiple verification strings; determines a mutation portion of the target string based on the probability distribution; and generates a mutant string by inputting a masked string in which the mutation portion of the target string is masked into the generative model and determining the character at the mutation portion based on the probability distribution obtained.

2. A computer system according to claim 1, wherein the processor calculates the maximum value of the probability distribution, and determines the mutation location based on the maximum value.

3. A computer system according to claim 2, wherein the processor receives constraint information regarding possible mutation locations in the target string, and generates a plurality of the verification strings based on the constraint information.

4. A computer system as described in claim 2, characterized in that the processor accepts mutation condition information regarding the number of mutation locations in the target string, and determines the mutation locations in the target string based on the mutation condition information and the probability distribution.

5. A computer system as described in claim 2, wherein the storage device stores information about a predictive model that predicts an index that evaluates the content represented by a string, and the processor performs reinforcement learning, including a process of generating a mutant string for training based on the probability distribution obtained by inputting the verification string generated from a training array into the generative model, a process of inputting the mutant string for training into the predictive model and obtaining the index, and a process of updating the generative model to improve the index.

6. A computer system according to claim 5, characterized in that the processor repeatedly executes the process of performing the reinforcement learning, the process of determining the mutation location of the target string, and the process of generating the mutated string.

7. A string generation method executed by a computer system, the computer system having a processor and a storage device connected to the processor, the storage device storing information about a generative model that accepts as input a string with at least one masked portion and outputs a probability distribution for each character in the masked portion, the string generation method comprising: a step by the processor accepting input of a target string; a step by the processor generating a plurality of verification strings in which one portion of the target string is masked; a step by the processor inputting the plurality of verification strings into the generative model and obtaining the probability distribution for each of the plurality of verification strings; a step by the processor determining a mutation portion of the target string based on the probability distribution; and a step by the processor generating a mutant string by inputting a masked string in which the mutation portion of the target string is masked into the generative model and determining the character at the mutation portion based on the probability distribution obtained.

8. A string generation method as described in claim 7, wherein the storage device stores information about a predictive model that predicts an index that evaluates the content represented by a string, and the string generation method is characterized in that it includes a step of performing reinforcement learning, including a process in which the processor generates a mutant string for training based on the probability distribution obtained by inputting the verification string generated from a training array into the generation model, a process in which the mutant string for training is input into the prediction model to obtain the index, and a process in which the generation model is updated to improve the index.

9. A computer system that converts at least one character of a string consisting of multiple characters and generates a new string, comprising: a processor; and a storage device connected to the processor, wherein the storage device stores information about a generative model that accepts as input a string with at least one masked portion and outputs a probability distribution of each character in the masked portion, and information about a mutation portion determination model that determines a mutation portion of the input string, wherein the processor: accepts input of a target string; inputs the target string into the mutation portion determination model to determine the mutation portion of the target string; generates a mutated string by determining the character at the mutation portion based on the probability distribution obtained by inputting a masked string in which the mutation portion of the target string is masked into the generative model; obtains an evaluation index for the mutated string; and learns the mutation portion determination model based on a reinforcement learning algorithm that uses the evaluation index of the mutated string as a reward.

Images