Array information processing device and method

Abstract

To support efficient creation of various array information in which elements constituting themselves secure variety.SOLUTION: An array information processing device 1 for processing array information indicated by array of a plurality of elements has: an attention analysis processing part 14 for identifying a pair between elements of array information in which contribution to output of a reward function for quantifying characteristics of the array information satisfies a predetermined condition, on the basis of the reward function; and a sampling function identification processing part 15 for calculating a sampling function for creating the array information so that a frequency of appearance of the plurality of elements becomes substantially equal in the identified pair between the elements.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to a technique for supporting the creation of sequence information that indicates the sequences of various elements such as amino acids. [Background technology]

[0002] There is a demand for technology that can efficiently create sequence information that shows the permutations and combinations of a large number of elements. For example, in the design of antibody drugs, multi-objective optimization is required to evaluate multifaceted properties such as affinity to antigens, hydrophobicity, solubility, and immunogenicity by combining amino acid sequences. Finding a sequence that simultaneously satisfies such multifaceted properties requires an enormous amount of calculation. For example, when there are 10 sequences and 20 amino acids, the total number of combinations is 20. 10 (>10 13 ), so a brute force approach to solving it is not realistic.

[0003] Therefore, Patent Document 1 discloses a technology for "automatically creating a large number of mutant sequences by mutating a part of the amino acid sequence of a prototype antibody." Patent Document 1 discloses that "physical properties of a prototype antibody and activity when the prototype antibody is combined with a predetermined target antigen are obtained as teacher data, and a computer is made to execute a process of generating a learning model in which the obtained physical properties are used as input data and the obtained activity is used as output data. The information processing method discloses that a computer is made to execute a process of obtaining an amino acid sequence of a mutant antibody different from the prototype antibody, obtaining the physical properties of the mutant antibody based on the obtained amino acid sequence of the mutant antibody, inputting the obtained physical properties of the mutant antibody into the learning model, and outputting the activity of the mutant antibody." [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2020-077206 A Summary of the Invention [Problem to be solved by the invention]

[0005] As described above, in Patent Document 1, a prediction model is constructed from experimental data on the properties and activity of mutant sequences, and candidates are extracted by applying reinforcement learning using highly active sequences and docking simulation. Here, in the background art including Patent Document 1, sequence information is treated as character string data, and characters are selected sequentially by Bayesian optimization or the like. The sequence information is expressed by a reward function indicating the characteristics as sequence information, and the sequence information created by this reward function is evaluated. In addition, by learning a sampling function proportional to the reward function based on flow matching conditions, sequence information for sequentially selecting characters based on the sampling function is created.

[0006] Here, regarding the reward function, there is a method of constructing a predictive model from the measured data by machine learning, etc. In this case, the accuracy of data that deviates from the measured data is degraded. In addition, the construction of a reward function based on the theoretical mechanisms of biology and chemistry is considered, but the accuracy may be insufficient because it simplifies complex biological phenomena. Due to the lack of performance of this reward function, when an actual experiment is performed on the created sequence, it often does not satisfy multifaceted characteristics. If the created sequence information is only similar, the majority of the characteristics may not be appropriate in actual experiments and demonstrations, and it may not be usable (for example, all may be annihilated). To avoid this, it is desirable to create as diverse sequence information as possible while satisfying the conditions as a drug for physical properties and activity based on the reward function. For example, it is required to create sequence information evenly without bias in the sequence space.

[0007] Therefore, an object of the present invention is to support the efficient creation of diverse sequence information while satisfying multifaceted characteristics based on a reward function. [Means for solving the problem]

[0008] In order to solve the above problems, in the present invention, in order to ensure the diversity of sequence information, an element pair of a sequence whose contribution to the output of a reward function indicating the characteristics of the sequence indicated by the sequence information satisfies a predetermined condition is identified. More specifically, in an array information processing device that processes sequence information indicated by an array of a plurality of elements, the array information processing device has an attention analysis processing unit that identifies an element pair of sequence information whose contribution to the output of the reward function satisfies a predetermined condition based on a reward function that quantifies the characteristics of the sequence information, and a sampling function identification processing unit that calculates a sampling function for creating the sequence information so that the occurrence frequency of the plurality of elements is approximately equal in the identified element pair.

[0009] The present invention also includes an array information processing device that processes array information represented by an array of a plurality of elements, the array information processing device having an input unit that accepts the plurality of elements as input, an attention analysis processing unit that identifies inter-element pairs of array information whose contribution to the output of a reward function satisfies a predetermined condition based on a reward function that quantifies a characteristic of the array information, a sampling function identification processing unit that calculates a sampling function for creating the array information so that the occurrence frequency of the plurality of elements is approximately equal in the identified inter-element pairs, and an output unit that outputs the created array information.

[0010] The present invention also includes an array information processing system having at least an array information processing device, and an array information processing method executed by the array information imaging device or the array information processing system.

[0011] Furthermore, the present invention also includes a program for causing the above-mentioned sequence information processing device to function as a computer and a storage medium for storing the program. Effect of the Invention

[0012] According to the present invention, it is possible to support the efficient creation of diverse sequence information in which the elements that constitute the sequence information ensure diversity. [Brief description of the drawings]

[0013] [Figure 1] FIG. 1 illustrates amino acids in one embodiment of the present invention. [Diagram 2] 1 is a system configuration diagram showing a sequence information processing system according to one embodiment of the present invention. [Diagram 3] FIG. 1 is a functional block diagram of an array information processing device 1 according to one embodiment of the present invention. [Figure 4] 1 is a hardware configuration diagram of a sequence information processing device 1 according to one embodiment of the present invention. [Diagram 5] FIG. 1 shows sequence information 18 used in one embodiment of the present invention. [Figure 6] FIG. 1 is a diagram for explaining the concept of one embodiment of the present invention. [Figure 7] 1 is a flowchart showing a process flow in one embodiment of the present invention. [Figure 8] FIG. 1 is a diagram for explaining attention in one embodiment of the present invention. [Figure 9] FIG. 11 is a diagram for explaining the degree of influence in one embodiment of the present invention. [Figure 10] 13 is a flowchart showing the sampling function learning process in step S7 in one embodiment of the present invention. [Figure 11] FIG. 13 is a diagram illustrating the strength of contribution of adjacent pairs in one embodiment of the present invention. [Figure 12] FIG. 2 is a diagram illustrating an evaluation index p of an adjacent pair in one embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] An embodiment of the present invention will be described below with reference to the drawings. In this embodiment, an antibody drug represented by an amino acid sequence having amino acids as elements will be described as an example. FIG. 1 is a diagram illustrating amino acids in this embodiment. In FIG. 1, a symbol is assigned to each amino acid (name). This symbol sequence (for example, sequence X=APRKNVRWCTISQPEWFKCR) represents the components of the antibody drug to be developed. Then, it is verified whether the antibody drug to be developed exerts a desired efficacy, which is an example of a characteristic. In this embodiment, an example of supporting the development of antibody drugs at multiple pharmaceutical companies will be described. Note that the symbols are an example of elements, and can also be expressed as characters (character strings / character patterns), etc.

[0015] Fig. 2 is a system configuration diagram showing a sequence information processing system in this embodiment. In Fig. 2, the sequence information processing system is realized as a cloud system that provides information services to a plurality of pharmaceutical companies. Specifically, in Fig. 2, each of the pharmaceutical companies 2a to 2c is connected to a sequence information processing device 1 realized as a server via a network 3. To achieve this, the pharmaceutical companies 2a to 2c have an in-house system having terminal devices 20a to 20c realized as PCs or the like, which is connected to the network 3.

[0016] Moreover, the terminal devices 20a to 20c are connected to each other via an in-house network such as an intranet. The terminal devices 20a to 20c notify the sequence information processing device 1 of various instructions and receive processing results from the sequence information processing device 1. For this reason, the terminal devices 20a to 20c each have a display device, an input device, a processing device, and a storage device, and the processing device executes various processes according to a program. In this embodiment, the sequence information processing device 1 is realized by an external server, but it may be realized as an in-house system like the sequence information processing device 1c in FIG. 2. Furthermore, the cloud system in this embodiment is one of the implementation examples, and the sequence information processing device 1 may be realized as a so-called standalone by a PC.

[0017] Next, the configuration of the sequence information processing device 1 that executes the main processing in this embodiment will be described. Fig. 3 is a functional block diagram of the sequence information processing device 1 in this embodiment. In Fig. 3, the sequence information processing device 1 has an input unit 11, an output unit 12, a reward function identification processing unit 13, an attention analysis processing unit 14, a sampling function identification processing unit 15, a sequence creation processing unit 16, and a storage unit 17.

[0018] Here, the input unit 11 accepts information, instructions, and the like from the terminal devices 20a to 20c, users, and the like. The input unit 11 is, for example, a keyboard, a mouse, a touch panel, a numeric keypad, a scanner, a microphone, and the like. The output unit 12 outputs information, processing results, and the like, for example, created array information, to the terminal devices 20a to 20c, users, and the like. The output unit 12 is, for example, a display, a printer, a speaker, and the like. The reward function identification processing unit 13 learns a prediction model to be predicted, and identifies, that is, specifies, it as a reward function (step S1). Here, the reward function quantifies the characteristics of the array information.

[0019] In addition, the attention analysis processing unit 14 aggregates the relationship (attention) between each amino acid, which is an element in the target sequence information, and identifies adjacent pairs (combinations) consisting of adjacent elements whose contribution to the reward function satisfies a predetermined condition at each position of the sequence information (step S2). In this embodiment, the predetermined condition is that the contribution to the reward function is smaller (weak) than a predetermined threshold. Note that the adjacent pair is an example of an inter-element pair consisting of a set of elements among the multiple elements that make up the sequence information. For this reason, an inter-element pair of elements at distant positions may be used.

[0020] Moreover, the sampling function identification processing unit 15 identifies a sampling function for creating the array information. Here, the sampling function is proportional to the reward function identified by the reward function identification processing unit 13 and is composed of the reward function. Here, the sampling function identification processing unit 15 identifies a sampling function that minimizes a loss function (loss function) using the adjacent pairs identified by the attention analysis processing unit 14 (step S3). Here, when "minimizing the loss function", it is desirable that the occurrence frequency of elements in the array information created by the sampling function is approximately uniform. Here, the occurrence frequency being approximately uniform means that the difference in the number of occurrences or frequency of each element in the array information is within a predetermined range.

[0021] Moreover, the array creation processing unit 16 creates array information using the sampling function identified by the sampling function identification processing unit 15. The creation of the array information may be performed by another device, or may be omitted in the array information processing device 1. The storage unit 17 stores array information 18 and evaluation index 19. Although not shown in FIG. 3, functions such as the above-mentioned reward function, loss function, and sampling function and rules for executing them may be stored. The array information 18 and evaluation index 19 will be described later.

[0022] Next, an implementation example of the sequence information processing device 1 will be described. Fig. 4 is a hardware configuration diagram of the sequence information processing device 1 in this embodiment. In Fig. 4, the sequence information processing device 1 has a processing device 101, a communication device 102, a memory 103, and a sub-storage device 104, which are connected to each other via a communication path.

[0023] First, the processing device 101 can be realized by a processor such as a CPU, and executes calculations according to an array information processing program 105 stored in a sub-storage device 104, which will be described later. The array information processing program 105 will be described later.

[0024] 3, is connected to the network 3, and communicates with other devices. In addition, in FIG. 3, the sequence information processing device 1 may be provided with an input device and a display device as the input unit 11 and the output unit 12.

[0025] 3. The memory 103 and the secondary storage device 104 expand the sequence information processing program 105 stored in the secondary storage device 104 and information used for processing by the processing device 101. The secondary storage device 104 can be realized by a so-called storage. The secondary storage device 104 stores the sequence information processing program 105, sequence information 18, and evaluation index 19. The secondary storage device 104 may be realized by various storage media such as an external HDD (Hard Disk Drive), SSD (Solid State Drive), or memory card, or may be realized by a device separate from the sequence information processing device 1, such as a file server.

[0026] Here, the sequence information processing program 105 is composed of a reward function identification processing module 106, an attention analysis processing module 107, a sampling function identification processing module 108, and a sequence creation processing module 109. Each of these modules may be realized by an individual program or a partial combination.

[0027] Here, the configuration shown in FIG. 3, which performs the same functions as each module, is as follows. Reward function identification processing unit 13: Reward function identification processing module 106 Attention analysis processing unit 14: Attention analysis processing module 107 Sampling function identification processing unit 15: Sampling function identification processing module 108 Sequence creation processing unit 16: Sequence creation processing module 109 Therefore, the processing device 101 executes the processes of the reward function identification processing unit 13, the attention analysis processing unit 14, the sampling function identification processing unit 15, and the array creation processing unit 16 in accordance with the array information processing program 105.

[0028] At least a part of the functions and processes of the array information processing device 1 may be provided in the terminal devices 20a to 20c.

[0029] Next, we will explain each piece of information used in this embodiment and stored in the secondary storage device 104. First, Fig. 5 is a diagram showing sequence information 18 used in this embodiment. The sequence information 18 is created by the sequence information processing device 1, and indicates the components of the antibody drug to be developed.

[0030] In Fig. 5, symbols 1 to N indicate amino acids that are elements. For example, the sequence information of NO1 is the sequence shown in Fig. 1, and is composed of the amino acids A: alanine, P: proline, R: arginine, K: lysine to R: arginine. In addition, in the sequence information of NO2, symbol 3 is changed to L: leucine compared to NO1. In this embodiment, it is assumed that M pieces of sequence information of NO1 to M are created.

[0031] Moreover, the evaluation index 19 is information used as teacher data for evaluating the array. The above is the end of the explanation of the information in this embodiment, and the following describes the processing in this embodiment. Here, first, the concept of this embodiment will be explained, and then the processing flow will be explained. FIG. 6 is a diagram for explaining the concept of this embodiment. In FIG. 6, (1) the array x is a combination of elements in the order of s'→s→s'', and a sampling function (F) that satisfies the flow matching condition when the elements are in this order is calculated. Here, as a flow matching condition, it is desirable that the inflow and outflow of elements are the same. However, it is very difficult or rare for them to always be "the same". Therefore, as shown in (Equation 1), a sampling function (F) that minimizes the loss function (L) as an approximate solution is obtained.

[0032]

number

[0033] Then, for adjacent pairs of adjacent elements whose contribution of the reward function satisfies a predetermined condition (for example, weaker than a threshold), a loss function (L) as shown in (Equation 2) is identified.

[0034]

number

[0035] Here, the significance of the above processing will be explained. First, in this embodiment, the aim is to randomize the arrangement of elements (level the arrangement space). However, simply randomizing the arrangement will result in an enormous amount of calculations, and there is no possibility of increasing the reward. For this reason, a sampling function that minimizes the loss function is used as an approximate solution for randomizing, that is, equalizing the probability, the "parts with low importance" for the characteristics such as efficacy. Here, the "parts with low importance" are those with a weak contribution to the reward function, that is, the contribution is preferably below a threshold, but it is sufficient that the contribution satisfies a predetermined condition.

[0036] 6, (2) sampling of adjacent pairs with weak contribution of reward function is performed to realize approximately equal probability of occurrence. As a result, the reward is shown large in the evaluation result of the function (closer to the desired efficacy, etc.), and randomized and diverse sequence information can be created.

[0037] Next, a process flow for realizing the above-mentioned concept will be described. Fig. 7 is a flowchart showing the process flow in this embodiment. In the following, this flowchart will be described using the configuration in Fig. 3.

[0038] First, in step S1, the reward function identification processing unit 13 acquires sequence information indicating the entire sequence x composed of amino acids constituting the antibody drug and a corresponding evaluation index p. For this purpose, the reward function identification processing unit 13 acquires amino acids accepted by the input unit 11, and creates an entire sequence x assumed by combining these. In addition, the reward function identification processing unit 13 may accept sequence information indicating the entire sequence x via the input unit 11, or may acquire sequence information indicating the entire sequence x stored in advance in the storage unit 17. Then, the reward function identification processing unit 13 acquires the corresponding evaluation index p from the storage unit 17. Note that the amino acid here is an example of an element.

[0039] In step S2, the reward function identification processing unit 13 performs encoding on each piece of acquired sequence information. For this purpose, for example, one-hot encoding is used.

[0040] In step S3, the reward function identification processing unit 13 uses the evaluation index p of each encoded sequence information as training data, and K , W V , W Q , β 1 p The following evaluation index p~β 1 p The following equations (3) to (9) show:

[0041]

number

[0042]

number

[0043]

number

[0044]

number

[0045]

number

[0046]

number

[0047]

number

[0048] In step S4, the reward function identification processing unit 13 uses the results of step S3 to determine the reward function R and the attention α t This reward function R is determined using (Equation 10).

[0049]

number

[0050] Also, in step S5, reward function identification processor 13 learns the reward function R identified in step S4. To this end, reward function identification processor 13 learns a prediction model that predicts characteristics such as affinity to an antigen, hydrophobicity, solubility, immunogenicity, etc. as reward function R. Here, the prediction model can also identify the relationship (attention) between amino acids (elements) in the sequence as the magnitude of contribution to the reward function R, in addition to the characteristic value, for each acquired sequence information.

[0051] Here, attention specified by the prediction model will be described. FIG. 8 is a diagram for explaining attention in this embodiment. In FIG. 8, X t indicates the sequence information, and each position (element) of the sequence information is indicated by x t n Also, the sequence information X t Impact level At The influence of each adjacent pair is α t nm Here, the sequence information X t These influences indicate the weighting of the importance of adjacent pairs during prediction processing using a prediction model. In particular, for antibody drugs, because they have a three-dimensional structure, amino acids at distant (non-adjacent) positions may have an influence on each other.

[0052] Next, the influence degree will be described. FIG. 9 is a diagram for explaining the influence degree in this embodiment. t Each position (element) x t i Adjacent position (element) of x t i+1 Taking the average of the attention of each position x all i,j From each amino acid, the influence of the adjacent position α all i,j Here, in FIG. 9, the thickness of the arrow between each element indicates the strength of influence. For this reason, the left side of FIG. 9 indicates that the influence of the adjacent pair A with any of A, C, and D is strong. And the right side of FIG. 9 indicates that the influence of the adjacent pair C with any of C, D, E, and F is strong.

[0053] In addition, the impact level α all i,j Depending on the influence of the reward function R(x), the contribution of the reward function R(x) can be treated as varying. For example, the influence of the all i,j If is small, the contribution to the reward function R(x) can also be treated as small. Below, we return to the explanation of the processing flow in Figure 7.

[0054] In addition, in step S6, the attention analysis processing unit 14 aggregates the relationships (attention) between amino acids in the entire sequence information acquired, and identifies adjacent pairs that have a weak contribution to the reward function R at each position of the sequence information. A weak contribution means that the contribution is equal to or less than a threshold value stored in advance in the storage unit 17. Also, "weak contribution" is just an example, and it is sufficient that the contribution of the adjacent pair satisfies a predetermined condition.

[0055] In step S7, the sampling function identification processing unit 15 performs learning to find a sampling function proportional to the reward function R from the sampling function (F) that minimizes the loss function.

[0056] For this purpose, Equation 11 is used. The details will be described below. Fig. 10 is a flowchart showing the sampling function learning process in step S7 in this embodiment. In step S71 in Fig. 10, the sampling function identification processor 15 selects learning data X for the array information.

[0057]

number

[0058] In step S72, the sampling function identification processing unit 15 selects an element s at a predetermined position from the acquired sequence information. In step S73, the sampling function identification processing unit 15 determines whether the contribution of the reward function R to the adjacent pair (s → s') including the selected element s is weak (below a threshold). If the contribution is weak (Y), the process proceeds to step S75. If the contribution is not weak (N), the process proceeds to step S74.

[0059] Here, the strength of the contribution of the adjacent pair in step S73 will be described with reference to FIG. 11 and FIG. 12. FIG. 11 is a diagram that shows the strength of the contribution of the adjacent pair in this embodiment. In FIG. 11, the ACDFs in the vertical direction each indicate an amino acid (symbol) that is an element of the sequence information, and by selecting this, sequence information showing an antibody drug in the horizontal direction is created. In the figure, the arrows connect the adjacent pairs, and the thickness of the arrows indicates the magnitude (strength) of the relationship between each amino acid. As in FIG. 9, FIG. 11 also shows that the influence of the adjacent pair between A and any of A, C, and D is strong. And FIG. 12 is a diagram that shows the evaluation index p of the adjacent pair in this embodiment. The evaluation index p for each of the elements s and s' to s''' is specified. In this way, it can be determined from FIG. 11 and FIG. 12 that an adjacent pair that has a weak contribution to the reward function R is an adjacent pair that has a small contribution to the reward function R.

[0060] Next, in step S74, the sampling function identification processor 15 calculates the loss function using Equation 12.

[0061]

number

[0062] Here, (Equation 12) represents an approximation of (Equation 13) which represents the flow matching condition.

[0063]

number

[0064] Furthermore, in step S75, the sampling function identification processing unit 15 calculates the loss function using Equation 14. In this manner, different loss functions are used in steps S74 and S75.

[0065]

number

[0066] Here, (Equation 14) has a regularization term that minimizes the entropy and minimizes the F(s→s′) value for adjacent pairs (s→s′) where the contribution of the reward function R is weak.

[0067] In step S76, the sampling function identification processing unit 15 judges whether the next position (element) exists in the sequence information for which the loss function should be calculated. If it exists (Y), the process proceeds to step S72, and the subsequent steps are executed for the next position. If it does not exist (N), the process proceeds to step S77.

[0068] Also, in step S77, the sampling function identification processor 15 judges whether there is next learning data X to be processed in this flow. As a result, if there is (Y), the process proceeds to step S71, and the subsequent steps are executed for the next learning data X. If there is not (N), the process ends, and the process proceeds to step S8 of step S7 in Fig. 7. This concludes the details of step S7, that is, the explanation of Fig. 10, and we return to the explanation of Fig. 7.

[0069] Next, in step S8, array creation processing unit 16 samples symbols, that is, a set of elements, using the sampling function (F) found in step S7, based on equation 11. As a result, array information is created.

[0070] This is the end of the processing flow in this embodiment, and the contents of this processing flow can be expressed as follows. An adjacent pair (s → s') that satisfies a predetermined condition, such as a weak contribution to the reward function R, is an adjacent pair that has a small contribution to the reward function R. In this embodiment, this small contribution is utilized to ensure diversity of sequence information. In minimizing the loss function, for adjacent pairs that have a weak contribution to the reward function R, the evaluation index for adjacent positions is (almost) equalized by entropy minimization. This equalization affects the achievement of the flow matching condition, but the regularization term reduces the value of F for the adjacent pair, minimizing the impact. In adjacent pairs that have a weak contribution to the reward function R, the selection of adjacent positions by the sampling function is randomized, resulting in a larger reward and a more diverse array of samples. Here, the reward indicates the degree of the desired characteristic or efficacy, and the larger the reward, the easier it is to obtain the desired characteristic or reward.

[0071] The reward can be expressed by reward items indicating affinity, hydrophobicity, solubility, immunogenicity, etc. The reward items may be used to screen out the learning data in step S71. That is, the sampling function identification processing unit 15 selectively acquires learning data (antibody drugs with sequence information) whose characteristic values ​​of reward items designated by a user or the like satisfy a predetermined criterion from among the reward items, and screens out the rest.

[0072] This concludes the description of this embodiment, but the present invention is not limited to this. For example, the elements are not limited to amino acids, and the sequence information is not limited to amino acid sequences. For example, the present invention can be applied to various chemical substances. Furthermore, the present invention can be applied to neodymium magnets, superconductors, semiconductors, and the like, which are created by changing the arrangement of atoms. Furthermore, the sequence information processing device 1 may also be realized by a computer other than a server. [Explanation of symbols]

[0073] 1...sequence information processing device, 11...input unit, 12...output unit, 13...reward function identification processing unit, 14...attention analysis processing unit, 15...sampling function identification processing unit, 16...sequence creation processing unit, 17...storage unit, 18...sequence information, 19...evaluation index, 101...processing device, 102...communication device, 103...memory, 104...sub-storage device, 2a to 2c...pharmaceutical company, 20a to 20c...terminal device, 3...network

Claims

1. In an array information processing device that processes array information represented by an array of a plurality of elements, an attention analysis processing unit that identifies an element pair of sequence information whose contribution to an output of the reward function satisfies a predetermined condition based on a reward function that quantifies a characteristic of the sequence information; a sampling function identification processing unit that calculates a sampling function for creating the array information such that the occurrence frequencies of the plurality of elements in the identified element pairs are approximately equal.

2. 2. The sequence information processing device according to claim 1, The array information processing device, wherein the predetermined condition indicates whether the contribution to the output of the reward function is less than a predetermined threshold.

3. 3. The array information processing device according to claim 2, The predetermined condition is an array information processing device that minimizes a loss function.

4. 2. The sequence information processing device according to claim 1, The sampling function identification processing unit is an array information processing device that selectively acquires learning data used to identify the sampling function by using reward items for the array information.

5. 2. The sequence information processing device according to claim 1, The sampling function identification processing unit uses loss functions that differ from one another depending on whether the predetermined condition is satisfied or not in identifying the sampling function.

6. 6. The array information processing device according to claim 5, When the predetermined condition is satisfied, a loss function having a regularization term that minimizes entropy is used.

7. 6. The array information processing device according to claim 5, If the predetermined condition is not satisfied, an approximation of the formula indicating the row matching condition is used.

8. 2. The sequence information processing device according to claim 1, The sequence information processing device further includes a reward function identification processing unit that identifies the reward function.

9. 2. The sequence information processing device according to claim 1, An array information processing device having an array creation processing unit that creates new array information by using the sampling function.

10. 10. The sequence information processing device according to claim 1, A sequence information processing device, wherein the elements are symbols representing amino acids, and the sequence information represents an antibody medical product composed of a plurality of amino acids.

11. In an array information processing device that processes array information represented by an array of a plurality of elements, an input unit that accepts the plurality of elements as an input; an attention analysis processing unit that identifies an element pair of sequence information whose contribution to an output of the reward function satisfies a predetermined condition based on a reward function that quantifies a characteristic of the sequence information; a sampling function identification processing unit that calculates a sampling function for creating the array information such that the occurrence frequencies of the plurality of elements in the identified element pairs are approximately equal; an output unit that outputs the generated sequence information; An array information processing device having the above structure.

12. An array information processing method executed by an array information processing device represented by an array of a plurality of elements, A process of identifying an element pair of sequence information whose contribution to the output of a reward function that quantifies a characteristic of the sequence information satisfies a predetermined condition based on the reward function; A process of calculating a sampling function for creating the sequence information such that the occurrence frequencies of the plurality of elements are approximately equal in the identified element pairs; A method for processing sequence information, comprising:

Images