Training program, inference program, training method, inference method, and information processing device

Abstract

This information processing device acquires training data in which input data and a label are associated with each other, the input data containing a plurality of primary structures included in a higher-order structure of a ligand-coupled receptor and structure information relating to the plurality of primary structures, the label indicating whether or not the receptor and the ligand can be coupled to each other. The information processing device executes machine learning of a machine learning model on the basis of the training data.

Description

Learning program, inference program, learning method, inference method, and information processing device

[0001] The present invention relates to a learning program and the like.

[0002] Receptors are regulatory proteins present in cells that selectively receive various signal molecules. They are mainly embedded in the cell membrane, but are also found in the cytoplasm and on the surface of the nucleus. The signal molecule that binds with a receptor to induce a biological response is called a "ligand."

[0003] Ligands include hormones, some amino acids, neurotransmitters, toxins, and pharmaceuticals. Ligands are known to selectively or specifically exhibit high affinity for specific sites on receptors. Different receptors often exist for different ligands, and the combinations of ligands and receptors that can bind vary greatly depending on the cell type.

[0004] Progress is being made in both protein conformational analysis and chemical characterization to infer possible receptor-ligand combinations.

[0005] The higher-order structures of proteins such as receptors are publicly available data and consist of sequences of approximately 20 types of amino acids. Figure 13A shows an example of the relationship between amino acid names, abbreviations, and symbols. For example, the abbreviation and symbol for the amino acid "alanine" are "Ala" and "A." The relationship between other amino acid names, abbreviations, and symbols is as shown in Figure 13A.

[0006] An amino acid is a molecule in which an amino group (-NH2) and a carboxy group (-COOH) are bonded to a carbon (C), and the general structural formula of an amino acid is shown in Figure 13B. Figure 13B is a diagram showing an example of the relationship between the general chemical structural formula of an amino acid and its side chain. A "side chain (R)" is also bonded to the central carbon (C), and the type of amino acid differs depending on this difference. Specific side chains (R) such as "alanine" and "valine" and their chemical structural formulas are shown below.

[0007] In conventional techniques, machine learning models are used to determine whether a combination of receptor and ligand is an appropriate combination (whether they can be combined).

[0008] FIG. 14 is a diagram (1) for explaining the conventional technology. In FIG. 14, the processing of the learning phase of the conventional technology is explained. For convenience of explanation, a device that executes the conventional technology will be referred to as a "conventional device." The conventional device executes machine learning of a machine learning model M1 using a pair of input data and a correct answer label.

[0009] For example, the input data may include multiple chemical structural formulas 5 of receptors and their respective atoms, as well as chemical structural formulas 6 of ligands to be combined with the receptors and their respective atoms. The receptor chemical structural formula 5 includes 5-1, 5-2, and 5-3. The correct answer label contains information indicating whether the receptors and ligands in the input data can be combined.

[0010] The conventional device uses a vector dictionary to calculate vectors vc5-1, vc5-2, and vc5-3 for multiple chemical structural formulas 5-1, 5-2, and 5-3 of the receptor based on the vectors of each atom. The conventional device uses a vector dictionary to calculate vector vc6 for chemical structural formula 6 based on the vectors of each atom. The conventional device calculates vector vc7 by multiplying vectors vc5-1, vc5-2, and vc5-3 with vector vc6.

[0011] The conventional device inputs the vector vc7 into the machine learning model M1 to obtain the output result 8. The conventional device updates the parameters of the machine learning model so that the difference between the output result 8 and the correct label becomes smaller.

[0012] The conventional device trains the machine learning model M1 by repeatedly executing the above process for other pairs of input data and correct labels.

[0013] Fig. 15 is a diagram (2) for explaining the conventional technique. Fig. 15 explains the processing of the inference phase of the conventional technique. The conventional device uses a trained machine learning model M1 to infer whether or not the receptor and ligand of the candidate data can be combined.

[0014] For example, the candidate data includes multiple chemical structural formulas 10 of receptors, each of their atoms, and chemical structural formulas 11 of ligands to be combined with the receptors. The chemical structural formulas 10 of the receptors include 10-1, 10-2, and 10-3.

[0015] The conventional device uses a vector dictionary to calculate vectors vc10-1, vc10-2, and vc10-3 of chemical structural formulas 10-1, 10-2, and 10-3 based on the vectors of each atom. The conventional device uses a vector dictionary to calculate vector vc11 of chemical structural formula 11 based on the vectors of each atom. The conventional device calculates vector vc12 by multiplying vectors vc10-1, vc10-2, and vc10-3 with vector vc11.

[0016] The conventional device inputs the vector vc12 into the trained machine learning model M1 to obtain an output result 13. If the output result 13 is "OK (combinable)," the conventional device estimates that the combination of the receptor and ligand in the candidate data is appropriate. On the other hand, if the output result 13 is "NG (not combinable)," the conventional device estimates that the combination of the receptor and ligand in the candidate data is inappropriate.

[0017] JP 2019-028879 A U.S. Patent Application Publication No. 2022 / 0246233 JP 2018-503171 A U.S. Patent Application Publication No. 2017 / 0323049 A

[0018] For example, whether a receptor and a ligand can combine is influenced not only by protein sequence information for chemical property analysis but also by coordinate information of the atoms that make up the receptor and ligand for three-dimensional structure analysis. However, although receptors are composed of higher-order protein structures (multiple primary structures) and ligands are composed of primary protein structures, the above-mentioned conventional technology focuses on each atom in the chemical structural formula of amino acids, which does not provide optimal granularity and information volume for estimation, resulting in the problem that it is not possible to properly estimate whether the target receptor and ligand can combine.

[0019] In one aspect, the present invention aims to provide a learning program, an inference program, a learning method, an inference method, and an information processing device that can appropriately estimate whether a target receptor and a ligand can combine.

[0020] In the first proposal, a computer is caused to execute the following process: The computer acquires training data in which input data having multiple primary structures included in the higher-order structure of a receptor combined with a ligand and structural information on the multiple primary structures is associated with a label indicating whether the receptor and ligand can combine. The computer executes machine learning of a machine learning model based on the training data.

[0021] It is possible to appropriately predict whether or not the target receptor and ligand can combine.

[0022] FIG. 1 is a diagram showing an example of protein structure data. FIG. 2 is a diagram explaining the processing of the preprocessing phase. FIG. 3 is a diagram showing an example of a Postscript program. FIG. 4 is a diagram (1) explaining the learning phase. FIG. 5 is a diagram (2) explaining the learning phase. FIG. 6 is a diagram (1) explaining the processing of the inference phase. FIG. 7 is a diagram (2) explaining the processing of the inference phase. FIG. 8 is a functional block diagram showing the configuration of an information processing device according to this embodiment. FIG. 9 is a flowchart showing the processing procedure of the preprocessing phase. FIG. 10 is a flowchart showing the processing procedure of the learning phase. FIG. 11 is a flowchart showing the processing procedure of the inference phase. FIG. 12 is a diagram showing an example of the hardware configuration of a computer that realizes the same functions as the information processing device of the embodiment. FIG. 13A is a diagram showing an example of the relationship between the names and symbols of amino acids. FIG. 13B is a diagram showing an example of the relationship between the general structural formula of the chemical structural formula of an amino acid and its side chain. FIG. 14 is a diagram (1) explaining the prior art. FIG. 15 is a diagram (2) explaining the prior art.

[0023] Hereinafter, embodiments of the learning program, inference program, learning method, inference method, and information processing device disclosed in the present application will be described in detail with reference to the accompanying drawings. However, the present invention is not limited to these embodiments.

[0024] Before describing the processing of the information processing device according to this embodiment, an example of "protein structure data" handled by the information processing device will be described. Protein structure data can be obtained from the Protein Data Bank (PDB).

[0025] 1 is a diagram showing an example of protein structure data. For example, the protein structure data 30 shown in FIG. 1 includes, as protein structure information, a header region 30a, a sequence information region 30b, and a coordinate information region 30c. The header region 30a contains information such as the molecular name corresponding to the protein.

[0026] The sequence information area 30b contains information about the sequence of amino acids contained in the protein. As described in FIG. 13A, the amino acid sequence information is information that lists the abbreviations (three letters) of the amino acids that make up the protein.

[0027] A sequence of multiple amino acids in the protein structure data corresponds to the primary structure of the protein. There are various patterns of amino acid sequences contained in the primary structure, but in this embodiment, the amino acid sequence contained in each primary structure is predefined. A sequence of multiple consecutive primary structures corresponds to the higher-order structure of the protein. In addition, the PDB may store data in which the receptor and ligand are combined with sequence information and coordinate information. For example, the protein structure data 30 includes the primary structures of multiple proteins that make up the higher-order structure of the receptor and the primary structure of the protein that makes up the ligand.

[0028] The coordinate information region 30c is set with the positions (three-dimensional coordinates) of multiple atoms that make up the amino acids contained in the protein. In this embodiment, the position of each atom that makes up the amino acids contained in the sequence information region 30b is set in the coordinate information region 30c. In this embodiment, attention is focused on a specific atom among the multiple atoms. For example, the specific atom may be (1) an amino group "N", (2) an atom located at the end of the side chain of the central carbon (C) (for example, the atom "C" in the amino acid valine "Val"), or (3) a carboxyl group "O". In the following description, the specific atoms of interest are referred to as the first atom, second atom, and third atom. There may be multiple first atoms in one primary structure, and the positions of each first atom may differ. The same applies to the second atom and the third atom.

[0029] As described above, by utilizing protein structural data of the receptor combined with the ligand, it is possible to identify the higher-order structure of the receptor (multiple primary structures), the one-dimensional structure of the ligand, and the coordinate information of specific atoms of the amino acids that make up the receptor and ligand.

[0030] Next, the processing of the information processing device according to this embodiment will be described. The information processing device according to this embodiment sequentially executes processing in a preprocessing phase, processing in a learning phase, and processing in an inference phase. In the following description, the information processing device according to this embodiment will be referred to as an "information processing device 100."

[0031] The following describes the processing of the pre-processing phase executed by the information processing device 100. Fig. 2 is a diagram for explaining the processing of the pre-processing phase. By executing the processing of the pre-processing phase, the information processing device 100 generates a first vector dictionary 142a and a second vector dictionary 142b.

[0032] The information processing device 100 has a protein structure database PDB 141. The protein structure database PDB 141 stores protein structure data corresponding to a plurality of proteins (receptors, or receptors combined with ligands). The protein structure data has been described with reference to FIG. 1.

[0033] The process of generating the first vector dictionary 142a by the information processing device 100 will be described. The information processing device 100 extracts multiple primary structures from each protein structure data in the protein structure database PDB 141. As described in FIG. 1, the information on the primary structures is stored in the sequence information area 30b.

[0034] In the explanation of Fig. 2, a plurality of primary structures (sequence information) will be collectively referred to as a "primary structure 41." As explained in Fig. 1, the primary structure is information on a character string of an amino acid sequence. The information processing device 100 breaks it down into character strings of amino acid sequences.

[0035] The information processing device 100 decomposes the primary structure 41 into character strings of multiple amino acid sequences (or functional group sequences of organic compounds), and then arranges the character strings of each primary structure in order. The information processing device 100 applies algorithms such as CBoW or skip-gram (Word2vec) to each arranged character string, treating each amino acid (or functional group) as a word, and calculates a vector for each character string of the primary structure corresponding to a sentence. The information processing device 100 registers the relationship between the character string of the reference unit of the primary structure 41 and the vector in the first vector dictionary 142a. Note that the information processing device 100 may divide the primary structure 41 into predefined reference units.

[0036] The information processing device 100 repeatedly executes the above process for other primary structures, thereby registering the relationship between the character strings of the amino acid sequences included in the other primary structures and the vectors in the first vector dictionary 142a. Note that the information processing device may assign vectors using amino acid units as reference units.

[0037] Next, a process by which the information processing device 100 generates the second vector dictionary 142b will be described. The information processing device 100 extracts coordinate information of multiple primary structures from each protein three-dimensional structure data in the protein structure database PDB 141. As described in FIG. 1, the coordinate information of the primary structure is information stored in the coordinate information area 30c, and includes information on the positions of the first atom, second atom, and third atom of each amino acid. The coordinate information is an example of "structural information."

[0038] The information processing device 100 generates a Postscript program character string that draws a three-dimensional line shape connecting the positions of a first atom, a second atom, and a third atom included in the coordinate information. FIG. 3 is a diagram showing an example of the Postscript program. For example, the example shown in FIG. 3 shows a three-dimensional line 50 connecting a first atom 50a1, a second atom 50a2, and a third atom 50a3 of the amino acid valine "Val." As shown in FIG. 3, the same atom may exist at multiple positions. For example, the information processing device 100 may generate the three-dimensional line 50 by repeatedly connecting the closest atoms among the multiple atoms.

[0039] The information processing device 100 generates a Postscript program 51 that draws a three-dimensional line 50. The Postscript program 51 includes a command statement (character string) for drawing the line 50. The information processing device 100 may also generate a Postscript program that projects a three-dimensional line onto a two-dimensional plane from a predetermined direction and draws the line projected onto the two-dimensional plane.

[0040] Returning to the description of FIG. 2 , the information processing apparatus 100 performs morphological analysis on the Postscript program to decompose it into multiple morphemes (tokens). After decomposing the Postscript program into multiple tokens, the information processing apparatus 100 arranges each token in order. The information processing apparatus 100 applies a CBoW or skip-gram algorithm to each of the arranged tokens, treats each token as a word, and calculates a vector for each token. The information processing apparatus 100 registers the relationship between the tokens of the Postscript program and the vectors in the second vector dictionary 142b.

[0041] The information processing device 100 repeatedly performs the above process for the coordinate information of other primary structures, thereby registering the relationship between the tokens of the Postscript program obtained from the coordinate information of other primary structures and the vectors in the second vector dictionary 142b.

[0042] As described above, the information processing device 100 generates the first vector dictionary 142a and the second vector dictionary 142b by executing the processing of the preprocessing phase. Note that the information processing device 100 may acquire in advance from an external device, etc., the first vector dictionary 142a, which defines the relationship between character strings of amino acid sequences and vectors, and the second vector dictionary 142b, which defines the relationship between tokens of Postscript programs and vectors. If the information processing device 100 acquires the first vector dictionary 142a and the second vector dictionary 142b in advance, it may skip the processing of the preprocessing phase.

[0043] Here, we demonstrate how to create a vector dictionary by treating the amino acid sequence of a protein's primary structure as a text sentence and each amino acid symbol as a word. Hiragana characters such as "te," "ni," "wo," and "wa" in Japanese are meaningful words. Approximately 20 amino acids also have chemical properties, such as acidic, basic, neutral / hydrophilic, and neutral / hydrophobic. Furthermore, words composed of multiple hiragana characters, such as "ai (love)" and "ai (indigo)," also have unique meanings. Therefore, adding amino acid sequences known as motifs that form regular three-dimensional structures such as protein alpha helices and beta sheets to a vector dictionary can improve the accuracy of the vector dictionary.

[0044] So far, we have demonstrated how to create a vector dictionary for biopharmaceutical ligands composed of amino acid sequences. Meanwhile, for organic compound pharmaceuticals composed of conventional functional group sequences, chemical property analysis and three-dimensional structural analysis can be performed by creating a vector dictionary calculated in the same way as for amino acid sequences, based on several dozen functional groups. Note that a method of assigning letters A to Z and symbols such as "!" to functional groups, similar to the assignment of letters A to Z to approximately 20 amino acids, can also be applied.

[0045] Next, a description will be given of the learning phase process executed by the information processing device 100. Figures 4 and 5 are diagrams for explaining the learning phase. The information processing device 100 executes the learning phase process using a teacher data table 143 prepared in advance.

[0046] Referring to FIG. 4, for example, the teacher data table 143 associates an item number, sequence information, coordinate information, and a label. The item number is a number that identifies a record (teacher data) in the teacher data table 143. The sequence information is the higher-order structure of the receptor combined with the ligand, and this higher-order structure includes a series of multiple primary structures. Furthermore, the item number can identify which primary structures that constitute the higher-order structure of the receptor are adjacent to the primary structure of the ligand in the combined state. The coordinate information is information that indicates the positions of the first atom, second atom, and third atom of each amino acid in multiple primary structures included in the higher-order structure of the protein. Note that one piece of coordinate information is set for one primary structure. The label indicates whether the receptor combined with the ligand is appropriate. For example, if the receptor combined with the ligand is appropriate, the label is set to "OK (e.g., 1)." On the other hand, if the receptor combined with the ligand is inappropriate, the label is set to "NG (e.g., 0)."

[0047] For example, the sequence information of item (1) includes, from the beginning, primary structures c1-1, c1-2, c1-3, and c1-4. For example, of the primary structures c1-1, c1-2, c1-3, and c1-4, the primary structure c1-3 is set as the primary structure of the ligand.

[0048] The coordinate information of item number (1) includes, from the top, coordinate information e1-1, e1-2, e1-3, and e1-4. For example, coordinate information e1-1 is information on the positions of the first atom, the second atom, and the third atom contained in primary structure c1-1. Coordinate information e1-2 is information on the positions of the first atom, the second atom, and the third atom contained in primary structure c1-2. Coordinate information e1-3 is information on the positions of the first atom, the second atom, and the third atom contained in primary structure c1-3. Coordinate information e1-4 is information on the positions of the first atom, the second atom, and the third atom contained in primary structure c1-4.

[0049] Here, the information processing device 100 calculates vectors of each primary structure for chemical property analysis using the first vector dictionary 142a created in the preparation phase. In the following description, the vectors of the primary structures are referred to as "sequence property vectors."

[0050] This will be explained using the primary structure c1-1 included in the sequence information of item number (1). The information processing device 100 breaks down the primary structure c1-1 into character strings in reference units (for example, predetermined atomic units, or predetermined atomic units). The information processing device 100 identifies the vector of each character string in the reference units of the primary structure c1-1 by comparing it with the first vector dictionary 142a. The information processing device 100 calculates the sequence characteristic vector vc1-1 of the primary structure c1-1 by integrating the vectors of each character string in the reference units.

[0051] The information processing device 100 calculates an array characteristic vector vc1-2 for the primary structure c1-2 in the same manner as for the primary structure c1-1. The information processing device 100 calculates an array characteristic vector vc1-3 for the primary structure c1-3. The information processing device 100 calculates an array characteristic vector vc1-4 for the primary structure c1-4.

[0052] Next, the information processing apparatus 100 calculates vectors for each piece of coordinate information using the second vector dictionary 142b created in the preparation phase. In the following description, the vectors for the coordinate information are referred to as "stereoscopic coordinate vectors."

[0053] The following description will be given using coordinate information e1-1 included in the coordinate information of item number (1). The information processing device 100 generates a Postscript program character string p1-1 that draws a three-dimensional line shape connecting the positions of the first atom, the second atom, and the third atom included in the coordinate information e1-1. The information processing device 100 performs morphological analysis on the Postscript program p1-1 to break it down into a plurality of morphemes (tokens).

[0054] The information processing device 100 identifies the vector of each token by comparing each token of the Postscript program p1-1 with the second vector dictionary 142b. The information processing device 100 calculates the three-dimensional coordinate vector vp1-1 by integrating the vectors of each token of the Postscript program p1-1.

[0055] The information processing apparatus 100 generates a Postscript program p1-2 for coordinate information e1-2 in the same manner as for coordinate information e1-1, and calculates a three-dimensional coordinate vector vp1-2 based on the Postscript program p1-2 and the second vector dictionary 142b. The information processing apparatus 100 generates a Postscript program p1-3 for coordinate information e1-3, and calculates a three-dimensional coordinate vector vp1-3 based on the Postscript program p1-3 and the second vector dictionary 142b. The information processing apparatus 100 generates a Postscript program p1-4 for coordinate information e1-4, and calculates a three-dimensional coordinate vector vp1-4 based on the Postscript program p1-4 and the second vector dictionary 142b.

[0056] 5. The information processing device 100 sequentially inputs pairs of sequence characteristic vectors of the sequence information of each primary structure of each item number and 3D coordinate vectors of the coordinate information into the machine learning model M1, and trains the machine learning model M1 (updates parameters) so that the values ​​output from the machine learning model M1 approach the corresponding labels.

[0057] The machine learning model M1 is a neural network (NN) such as BERT (Pre-training of Deep Bidirectional Transformers for Language Understanding), Next Sentence Prediction, or Transformers.

[0058] A case will be described in which the information processing device 100 updates parameters using the array information, coordinate information, and label corresponding to item number (1) in the teacher data table 143.

[0059] 4, the information processing device 100 calculates sequence characteristic vectors vc1-1 to vc1-4 of the primary structures of the sequence information corresponding to item number (1). The information processing device 100 also calculates three-dimensional coordinate vectors vp1-1 to vp1-4 of the coordinate information corresponding to item number (1).

[0060] The information processing device 100 sequentially inputs pairs of array characteristic vectors of array information and cubic coordinate vectors of coordinate information to the machine learning model M1. For example, the information processing device 100 first inputs the array characteristic vector vc1-1 and the cubic coordinate vector vp1-1 to the machine learning model M1. The information processing device 100 secondly inputs the array characteristic vector vc1-2 and the cubic coordinate vector vp1-2 to the machine learning model M1. The information processing device 100 thirdly inputs the array characteristic vector vc1-3 and the cubic coordinate vector vp1-3 to the machine learning model M1. The information processing device 100 fourthly inputs the array characteristic vector vc1-4 and the cubic coordinate vector vp1-4 to the machine learning model M1.

[0061] When the information processing device 100 inputs a pair of the sequence characteristic vector of the last primary structure and the three-dimensional coordinate vector of the coordinate information into the machine learning model M1, it updates the parameters of the machine learning model M1 so that the difference between the output result output from the machine learning model M1 and the label of item number (1) becomes small.

[0062] The information processing device 100 updates the parameters of the machine learning model M1 by repeatedly performing the same process as above for the array information, coordinate information, and labels from item number (2) onwards in the teacher data table 143.

[0063] Next, the processing of the inference phase executed by the information processing device 100 will be described. Figures 6 and 7 are diagrams for explaining the processing of the inference phase. First, Figure 6 will be described. The information processing device 100 receives, from a user, sequence information and coordinate information of a certain receptor that is an inference target and that has combined with a certain ligand.

[0064] For example, the sequence information includes, from the beginning, primary structures c10-1, c10-2, c10-3, and c10-4. For example, of the primary structures c10-1, c10-2, c10-3, and c10-4, the primary structure c10-3 is set as the primary structure of the ligand.

[0065] The coordinate information includes, in order from the beginning, coordinate information e10-1, e10-2, e10-3, and e10-4. For example, coordinate information e10-1 is information on the positions of the first atom, the second atom, and the third atom contained in the primary structure c10-1. Coordinate information e10-2 is information on the positions of the first atom, the second atom, and the third atom contained in the primary structure c10-2. Coordinate information e10-3 is information on the positions of the first atom, the second atom, and the third atom contained in the primary structure c10-3. Coordinate information e10-4 is information on the positions of the first atom, the second atom, and the third atom contained in the primary structure c10-4.

[0066] Here, the information processing device 100 calculates sequence characteristic vectors vc10-1, vc10-2, vc10-3, and vc10-4 of the primary structures c10-1, c10-2, c10-3, and c10-4 using the first vector dictionary 142a. The process by which the information processing device 100 calculates the sequence characteristic vectors of the primary structures using the first vector dictionary 142a is similar to the process described in the learning phase.

[0067] Next, the information processing device 100 generates Postscript programs p10-1, p10-2, p10-3, and p10-4 based on the coordinate information e10-1, e10-2, e10-3, and e10-4. The process by which the information processing device 100 generates Postscript programs based on the coordinate information is similar to the process described in the learning phase.

[0068] The information processing device 100 calculates the three-dimensional coordinate vectors vp10-1, vp10-2, vp10-3, and vp10-4 of the Postscript programs p10-1, p10-2, p10-3, and p10-4 using the second vector dictionary 142b. The process by which the information processing device 100 calculates the three-dimensional coordinate vectors using the second vector dictionary 142b is similar to the process described in the learning phase.

[0069] 7, the information processing device 100 sequentially inputs the sets of array characteristic vectors vc10-1 to vc10-4 of the array information and the three-dimensional coordinate vectors vp10-1 to vp10-4 of the coordinate information, which were described with reference to FIG.

[0070] For example, the information processing device 100 first inputs the array characteristic vector vc10-1 and the cubic coordinate vector vp10-1 to the machine learning model M1. The information processing device 100 secondly inputs the array characteristic vector vc10-2 and the cubic coordinate vector vp10-2 to the machine learning model M1. The information processing device 100 thirdly inputs the array characteristic vector vc10-3 and the cubic coordinate vector vp10-3 to the machine learning model M1. The information processing device 100 fourthly inputs the array characteristic vector vc10-4 and the cubic coordinate vector vp10-4 to the machine learning model M1.

[0071] The information processing device 100 acquires an output result from the machine learning model M1 at the point when the pair of the sequence characteristic vector of the final primary structure and the 3D coordinate vector of the coordinate information is input to the machine learning model M1. For example, the machine learning model M1 in this embodiment may output a score indicating the likelihood of successful merging.

[0072] If the score of the output result is equal to or greater than the threshold (combination OK), the information processing device 100 estimates that the receptor shown in the sequence information in Figure 6 combined with the ligand is appropriate (the sequence of the receptor's primary structure is also appropriate). On the other hand, if the score of the output result is less than the threshold (combination NG), the information processing device 100 estimates that the receptor shown in the sequence information in Figure 6 combined with the ligand is inappropriate.

[0073] As described above, in the learning phase, the information processing device 100 performs machine learning of the machine learning model based on training data in which input data having sequence information of multiple primary structures included in the higher-order structure of the receptor combined with the ligand and coordinate information of the multiple primary structures is associated with labels indicating whether the receptor and the ligand can combine, thereby generating a machine learning model M1 for appropriately estimating whether the target receptor and ligand can combine.

[0074] The information processing device 100 inputs data containing sequence information of multiple primary structures contained in the higher-order structure of a receptor combined with a target ligand and coordinate information of the multiple primary structures into the trained machine learning model M1. The information processing device 100 can estimate whether the receptor combined with the ligand is appropriate by using the output results. Furthermore, in this embodiment, it is also possible to determine whether the arrangement of multiple primary structures of the receptor (including the primary structure of the ligand) is appropriate.

[0075] Next, an example of the configuration of an information processing device 100 that executes the above-mentioned inference phase process, learning phase process, and inference phase process will be described. Fig. 8 is a functional block diagram showing the configuration of an information processing device according to this embodiment. As shown in Fig. 8, the information processing device 100 includes a communication unit 110, an input unit 120, a display unit 130, a storage unit 140, and a control unit 150.

[0076] The communication unit 110 is connected to an external device or the like via a wired or wireless connection, and transmits and receives information to and from the external device or the like. The communication unit 110 is realized by a NIC (Network Interface Card) or the like. The communication unit 110 may be connected to a network (not shown). The communication unit 110 may receive information on the protein structure database PDB 141, the first vector dictionary 142 a, and the second vector dictionary 142 b from the external device, and register the information in the storage unit 140.

[0077] The input unit 120 is an input device that inputs various types of information to the information processing device 100. The input unit 120 corresponds to a keyboard, a mouse, a touch panel, etc. For example, in the inference phase, a user may operate the input unit 120 to input array information and coordinate information to be inferred.

[0078] The display unit 130 is a display device that displays information output from the control unit 150. The display unit 130 corresponds to a liquid crystal display, an organic EL (Electro Luminescence) display, a touch panel, etc. For example, the display unit 130 displays the estimation result of the estimation phase by the control unit 150.

[0079] The storage unit 140 includes a protein structure database PDB 141, a first vector dictionary 142 a, a second vector dictionary 142 b, a teacher data table 143, and a machine learning model M1. The storage unit 140 is realized by, for example, a semiconductor memory element such as a random access memory (RAM) or a flash memory, or a storage device such as a hard disk or an optical disk.

[0080] The protein structure database PDB141 stores protein structure data corresponding to a plurality of proteins (receptors, or receptors combined with ligands). The explanation of the protein structure database PDB141 is the same as that given in FIG.

[0081] The first vector dictionary 142a is a dictionary that stores character strings, which are basic units of primary structures, in association with vectors. Other descriptions of the first vector dictionary 142a are the same as those of the first vector dictionary 142a shown in FIG. 2 and the like.

[0082] The second vector dictionary 142b is a dictionary that stores tokens of a Postscript program generated from coordinate information in association with vectors. Other descriptions of the second vector dictionary 142b are the same as those of the second vector dictionary 142b shown in FIG. 2 and the like.

[0083] The teacher data table 143 holds multiple teacher data. The teacher data associates array information, coordinate information, and labels. Each teacher data is used when performing machine learning on the machine learning model M1. The explanation of the data structure of the teacher data table 143 is the same as the explanation of the data structure of the teacher data table 143 shown in FIG. 4.

[0084] The machine learning model M1 is a neural network such as BERT, Next Sentence Prediction, or Transformers, which are described in FIG. 5 .

[0085] The control unit 150 includes a preprocessing unit 151, a learning processing unit 152, and an inference processing unit 153. The control unit 150 is realized by, for example, a central processing unit (CPU) or a micro processing unit (MPU). The control unit 150 may also be implemented by an integrated circuit such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA).

[0086] The preprocessing unit 151 executes the processing of the preprocessing phase described with reference to Fig. 2 etc. The preprocessing unit 151 acquires protein structure data from the protein structure database PDB 141, and acquires from the protein structure data a plurality of primary structures and coordinate information corresponding to each primary structure.

[0087] The preprocessing unit 151 decomposes the primary structure into character strings of multiple reference units (e.g., amino acid sequences), applies algorithms such as CBoW or skip-gram (Word2vec), and assigns a vector to each character string. The preprocessing unit 151 sets the relationship between the character strings and the vectors in the first vector dictionary 142a.

[0088] The preprocessing unit 151 generates a Postscript program that draws the shape of a three-dimensional line connecting the positions of the first atom, the second atom, and the third atom included in the coordinate information. The preprocessing unit 151 performs morphological analysis on the Postscript program to decompose it into multiple morphemes (tokens). The preprocessing unit 151 applies algorithms such as CBoW or skip-gram (Word2vec) to assign a vector to each token, and sets the relationship between the tokens and vectors of the Postscript program in the second vector dictionary 142b.

[0089] The other processes of the preprocessing unit 151 are the same as those of the preprocessing phase described with reference to Figures 2 and 3. Note that the preprocessing unit 151 may skip the preprocessing phase if the first vector dictionary 142a and the second vector dictionary 142b have been acquired in advance.

[0090] The learning processing unit 152 executes the learning phase processing described in Figures 4 and 5. The learning processing unit 152 acquires training data from the training data table 143. The learning processing unit 152 calculates a sequence characteristic vector for each primary structure of the sequence information based on the first vector dictionary 142a. The learning processing unit 152 generates a Postscript program character string from multiple pieces of coordinate information included in the coordinate information, and calculates each three-dimensional coordinate vector based on the second vector dictionary 142b.

[0091] The learning processing unit 152 inputs the pair of the array characteristic vector and the cubic coordinate vector into the machine learning model M1, and updates the parameters of the machine learning model M1 based on the backpropagation method or the like so as to reduce the difference between the output result of the machine learning model M1 and the label. The learning processing unit 152 repeatedly executes the above process using each piece of training data stored in the training data table 143.

[0092] The other processes executed by the learning processing unit 152 are the same as those in the learning phase described with reference to FIGS.

[0093] The inference processing unit 153 executes the processing of the inference phase described with reference to Fig. 6 and Fig. 7. The inference processing unit 153 acquires array information and coordinate information to be inferred from an external device or the input unit 120.

[0094] The inference processing unit 153 calculates sequence characteristic vectors for each primary structure of the sequence information using the first vector dictionary 142a. The learning processing unit 152 generates a Postscript program from multiple pieces of coordinate information included in the coordinate information, and calculates each three-dimensional coordinate vector based on the second vector dictionary 142b.

[0095] The inference processing unit 153 inputs pairs of sequence characteristic vectors and 3D coordinate vectors to the machine learning model M1 in order from the beginning. When the inference processing unit 153 inputs the pair of the sequence characteristic vector of the last primary structure and the 3D coordinate vector of the coordinate information to the machine learning model M1, it acquires the output result output from the machine learning model M1.

[0096] If the score of the output result is equal to or greater than the threshold (combination OK), the inference processing unit 153 infers that the receptor indicated in the sequence information of the inference target and combined with the ligand is appropriate (the sequence of the receptor's primary structure is also appropriate).On the other hand, if the score of the output result is less than the threshold (combination NG), the inference processing unit 153 infers that the receptor indicated in the sequence information of the inference target and combined with the ligand is inappropriate.

[0097] The inference processing unit 153 outputs the inference result to the display unit 130 for display.

[0098] The other processes executed by the inference processing unit 153 are similar to the processes in the inference phase described with reference to FIGS.

[0099] Next, an example of the processing procedure of the information processing device 100 according to this embodiment will be described. Fig. 9 is a flowchart showing the processing procedure of the preprocessing phase. As shown in Fig. 9, the preprocessing unit 151 of the information processing device 100 acquires protein structure data from the protein structure database PDB 141 (step S101). The preprocessing unit 151 acquires multiple primary structures and coordinate information corresponding to each primary structure from the protein structure data (step S102).

[0100] The preprocessing unit 151 decomposes the primary structure into character strings of a plurality of reference units and assigns a vector to each character string (step S103). The preprocessing unit 151 sets the relationship between the character strings and the vectors in the first vector dictionary 142a (step S104).

[0101] The preprocessing unit 151 generates a Postscript program that draws the shape of a line connecting the positions of the first atom, the second atom, and the third atom included in the coordinate information (step S105). The preprocessing unit 151 decomposes the character string of the Postscript program into multiple tokens and assigns a vector to each token (step S106).

[0102] The preprocessing unit 151 sets the relationship between the token and the vector in the second vector dictionary 142b (step S107).

[0103] 10 is a flowchart showing the processing steps of the learning phase. As shown in FIG. 10, the learning processing unit 152 of the information processing device 100 acquires training data (sequence information and coordinate information) from the training data table 143 (step S201). The learning processing unit 152 calculates a sequence characteristic vector of each primary structure included in the sequence information based on the first vector dictionary 142a (step S202).

[0104] The learning processing unit 152 generates character strings for a plurality of Postscript programs from the plurality of pieces of coordinate information (step S203). The learning processing unit 152 calculates a three-dimensional coordinate vector from each Postscript program based on the second vector dictionary 142b (step S204).

[0105] The learning processing unit 152 inputs the pair of the array characteristic vector and the cubic coordinate vector to the machine learning model M1 (step S205). The learning processing unit 152 calculates the difference between the output result of the machine learning model M1 and the label (step S206).

[0106] The learning processing unit 152 updates the parameters of the machine learning model M1 so as to reduce the difference (step S207). If the learning processing unit 152 determines to continue the process (step S208, Yes), the learning processing unit 152 proceeds to step S201. On the other hand, if the learning processing unit 152 determines not to continue the process (step S208, No), the learning processing unit 152 ends the process.

[0107] 11 is a flowchart showing the processing procedure of the inference phase. As shown in FIG. 11, the inference processing unit 153 of the information processing device 100 acquires sequence information and coordinate information of the receptor and ligand to be inferred from the input unit 120 (step S301). The inference processing unit 153 calculates a sequence characteristic vector of each primary structure included in the sequence information based on the first vector dictionary 142a (step S302).

[0108] The inference processing unit 153 generates a plurality of Postscript programs from the plurality of pieces of coordinate information (step S303).The inference processing unit 153 calculates a three-dimensional coordinate vector from each Postscript program based on the second vector dictionary 142b (step S304).

[0109] The inference processing unit 153 inputs a pair of the sequence characteristic vector of the primary structure and the 3D coordinate vector into the machine learning model M1 (step S305). The inference processing unit 153 determines whether the target receptor (receptor combined with the ligand) is appropriate based on the output result of the machine learning model M1 (step S306). The inference processing unit 153 displays the determination result on the display unit 130 (step S307).

[0110] Next, the effects of the information processing device 100 according to this embodiment will be described. In the learning phase, the information processing device 100 trains the machine learning model M1 using training data that associates correct labels with input data having sequence information and coordinate information for multiple primary structures included in the higher-order structure of a receptor combined with a ligand. Furthermore, in the combined state, it is possible to identify which primary structures that make up the higher-order structure of the receptor are adjacent to the primary structure of the ligand. This makes it possible to generate a machine learning model M1 that appropriately estimates whether a target receptor and ligand can combine.

[0111] The information processing device 100 sequentially inputs pairs of primary structures and coordinate information corresponding to the primary structures into the machine learning model M1, and executes machine learning of the machine learning model M1 so as to reduce the difference between the output result of the machine learning model M1 and the label. This makes it possible to generate a machine learning model M1 that not only appropriately estimates whether a target receptor and a ligand can combine, but also estimates whether the sequence of the primary structure of the receptor combined with the ligand is appropriate.

[0112] The information processing device 100 generates coordinate information based on the positions of predetermined atoms contained in multiple primary structures, thereby generating a machine learning model M1 for estimating whether the sequence of the primary structure of a receptor combined with a ligand is appropriate, using both the sequence of the primary structure and the positions of the atoms.

[0113] The information processing device 100 generates coordinate information by converting a character string of a PostScript program that draws line segments connecting the positions of predetermined atoms contained in multiple primary structures into vectors. This allows the positions of the predetermined atoms to be handled as vectors, making it possible to efficiently perform machine learning on the machine learning model M1.

[0114] The information processing device 100 generates sequence information by converting the primary structure into a vector based on a dictionary that associates character strings of basic units of proteins with vectors. This allows the primary structure to be handled as a vector, and machine learning can be efficiently performed on the machine learning model M1.

[0115] In the inference phase, the information processing device 100 inputs input data containing sequence information and coordinate information of multiple primary structures contained in the higher-order structure of a target receptor (a receptor combined with a ligand) into a trained machine learning model M1, thereby making it possible to appropriately estimate whether the target receptor and ligand can combine.

[0116] The information processing device 100 sequentially inputs pairs of primary structures and coordinate information corresponding to the primary structures into the trained machine learning model M1, and obtains output results from the machine learning model M1. This not only makes it possible to appropriately estimate whether a target receptor and a ligand can combine, but also makes it possible to estimate whether the sequence of the primary structure of the receptor combined with the ligand is appropriate.

[0117] Next, an example of the hardware configuration of a computer that realizes the same functions as the information processing device 100 described in the above embodiment will be described. Fig. 12 is a diagram showing an example of the hardware configuration of a computer that realizes the same functions as the information processing device of the embodiment.

[0118] 12, computer 300 has a CPU 301 that executes various types of arithmetic processing, an input device 302 that accepts data input from a user, and a display 303. Computer 300 also has a communication device 304 that exchanges data with external devices via a wired or wireless network, and an interface device 305. Computer 300 also has a RAM 306 that temporarily stores various types of information, and a hard disk drive 307. Each of devices 301 to 307 is connected to a bus 308.

[0119] The hard disk drive 307 stores a pre-processing program 307 a, a learning processing program 307 b, and an inference processing program 307 c. The CPU 301 reads out each of the programs 307 a to 307 c and loads them into the RAM 306.

[0120] The pre-processing program 307a functions as the pre-processing process 306a, the learning processing program 307b functions as the learning processing process 306b, and the inference processing program 307c functions as the inference processing process 306c.

[0121] The pre-processing process 306a corresponds to the process by the pre-processing unit 151. The learning process 306b corresponds to the process by the learning processing unit 152. The inference process 306c corresponds to the process by the inference processing unit 153.

[0122] It should be noted that each of the programs 307a to 307c does not necessarily have to be stored in the hard disk drive 307 from the beginning. For example, each program may be stored in a "portable physical medium" such as a flexible disk (FD), CD-ROM, DVD, magneto-optical disk, or IC card that is inserted into the computer 300. Then, the computer 300 may read and execute each of the programs 307a to 307c.

[0123] REFERENCE SIGNS LIST 100 Information processing device 110 Communication unit 120 Input unit 130 Display unit 140 Storage unit 141 Protein structure DB 142a First vector dictionary 142b Second vector dictionary 143 Teacher data table 150 Control unit 151 Preprocessing unit 152 Learning processing unit 153 Inference processing unit

Claims

1. acquiring training data in which input data having a plurality of primary structures included in the higher-order structure of the receptor combined with the ligand and structural information of the plurality of primary structures is associated with a label indicating whether the receptor and the ligand can combine; Execute machine learning for the machine learning model based on the training data. A learning program that causes a computer to execute a process.

2. the higher-order structure of the input data includes a primary structure of the ligand and a plurality of primary structures other than the primary structure of the ligand; The learning program according to claim 1, characterized in that the process of performing the machine learning involves inputting pairs of the primary structure and structural information into the machine learning model in order, and performing machine learning of the machine learning model so that the difference between the output result of the machine learning model and the label becomes small.

3. 2. The learning program according to claim 1, further causing a computer to execute a process of generating the structural information based on the positions of predetermined atoms contained in the primary structure.

4. The learning program according to claim 2, further comprising causing a computer to convert a PostScript character string that draws a line segment connecting the positions of specified atoms contained in the primary structure into a vector.

5. The learning program according to claim 1, further comprising causing a computer to execute a process of converting the primary structure into a vector by dividing the primary structure into character strings of amino acid sequences of proteins or functional group sequences of organic compounds, and assigning a vector to each character or each character string.

6. a target higher-order structure of a target receptor to be inferred, and a plurality of target primary structures included in the target higher-order structure of the target receptor combined with a target ligand, and a plurality of target structure information corresponding to the plurality of target primary structures; The method infers whether the target receptor is suitable by inputting the target primary structures and the target structural information to a machine learning model that has been subjected to machine learning based on training data in which input data having a plurality of primary structures included in the higher-order structure of a receptor combined with a ligand and structural information of the plurality of primary structures is associated with a label indicating whether the receptor and the ligand can combine. An inference program that causes a computer to execute processing.

7. the target higher-order structure includes a target primary structure of the target ligand and a plurality of target primary structures other than the target primary structure of the target ligand; The inference program according to claim 6, characterized in that the inference process inputs pairs of the target primary structure and the target structural information into the machine learning model in order, and infers whether the target receptor is appropriate based on the output result of the machine learning model.

8. acquiring training data in which input data having a plurality of primary structures included in the higher-order structure of the receptor combined with the ligand and structural information of the plurality of primary structures is associated with a label indicating whether the receptor and the ligand can combine; Execute machine learning for the machine learning model based on the training data. A learning method characterized in that the processing is executed by a computer.

9. a target higher-order structure of a target receptor to be inferred, and a plurality of target primary structures included in the target higher-order structure of the target receptor combined with a target ligand, and a plurality of target structure information corresponding to the plurality of target primary structures; The method infers whether the target receptor is suitable by inputting the target primary structures and the target structural information to a machine learning model that has been subjected to machine learning based on training data in which input data having a plurality of primary structures included in the higher-order structure of a receptor combined with a ligand and structural information of the plurality of primary structures is associated with a label indicating whether the receptor and the ligand can combine. An inference method characterized in that the processing is executed by a computer.

10. acquiring training data in which input data having a plurality of primary structures included in the higher-order structure of the receptor combined with the ligand and structural information of the plurality of primary structures is associated with a label indicating whether the receptor and the ligand can combine; Execute machine learning for the machine learning model based on the training data. An information processing device having a control unit that executes processing.

11. a target higher-order structure of a target receptor to be inferred, and a plurality of target primary structures included in the target higher-order structure of the target receptor combined with a target ligand, and a plurality of target structure information corresponding to the plurality of target primary structures; The method infers whether the target receptor is suitable by inputting the target primary structures and the target structural information to a machine learning model that has been subjected to machine learning based on training data in which input data having a plurality of primary structures included in the higher-order structure of a receptor combined with a ligand and structural information of the plurality of primary structures is associated with a label indicating whether the receptor and the ligand can combine. An information processing device having a control unit that executes processing.