Computer program, information processing apparatus, information processing method, and learning model generation method

Abstract

To provide a computer program, an information processing device, an information processing method, and a learning model generation method capable of performing calculation of interaction between molecules faster than quantum chemical calculation.SOLUTION: The computer program causes a computer to execute a process of acquiring a feature value of a complex and a molecule contained in the complex, acquiring a molecular formation energy of the complex and the molecule contained in the complex output by a learning model by inputting the acquired feature value to the learning model that outputs the molecular formation energy of the complex and the molecule contained in the complex when the feature value of the complex and the molecule contained in the complex is input, and estimating an interaction energy between the molecules contained in the complex based on the acquired molecular formation energy.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a computer program, an information processing device, an information processing method, and a learning model generation method. [Background technology]

[0002] Drug discovery is the process of creating new medicines, and requires many steps, such as discovering and optimizing hit compounds that form the basis of drugs, conducting clinical trials, and applying for approval from the Ministry of Health, Labor and Welfare. Drug discovery is said to take more than 10 years and cost tens of billions of yen. Furthermore, to discover a hit compound, it is necessary to evaluate many drug compounds that interact with target proteins, and quantum chemical calculations, for example, are used to evaluate these interactions.

[0003] Quantum chemical calculation is a method for analyzing the structure and properties of atoms and molecules from their electronic states, and uses the Schrödinger equation. In actual calculations, some kind of approximate formula is introduced to solve the Schrödinger equation. Patent Document 1 discloses an information processing program that determines a quantum chemical calculation algorithm that can be executed within a specified time in order to understand molecular properties that will be useful in drug discovery and the discovery of new materials. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Laid-Open No. 2024-67991 Summary of the Invention [Problem to be solved by the invention]

[0005] Quantum chemical calculations can calculate intermolecular interactions with high precision, but they have the problem of taking a long time to complete the calculations (for example, it can take several tens of hours) and being expensive to perform.

[0006] The present invention has been made in view of the above circumstances, and aims to provide a computer program, an information processing device, an information processing method, and a learning model generation method that can calculate intermolecular interactions faster than quantum chemical calculations. [Means for solving the problem]

[0007] The present application includes multiple means for solving the above-mentioned problems. As an example, a computer program causes a computer to execute a process of acquiring features of a complex and molecules contained in the complex, inputting the acquired features into a learning model that outputs molecular formation energies of the complex and the molecules contained in the complex when the features of the complex and the molecules contained in the complex are input, acquiring the molecular formation energies of the complex and the molecules contained in the complex output by the learning model, and estimating the interaction energy between the molecules contained in the complex based on the acquired molecular formation energies. [Effects of the Invention]

[0008] According to the present invention, it is possible to calculate intermolecular interactions at a higher speed than quantum chemical calculations. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 illustrates an example of a configuration of an information processing device. [Figure 2] FIG. 1 is a diagram showing an example of information recorded in a polymolecular system structure DB. [Figure 3] FIG. 1 is a diagram showing an example of information recorded in a single-molecule structure DB. [Figure 4] FIG. 1 is a diagram illustrating a first example of a method for generating a learning model. [Figure 5] FIG. 10 is a diagram illustrating a second example of a method for generating a learning model. [Figure 6] FIG. 1 is a diagram showing a first example of a method for estimating intermolecular interaction energy using an information processing device. [Figure 7] FIG. 10 is a diagram showing a second example of a method for estimating intermolecular interaction energy using an information processing device. [Figure 8] FIG. 10 is a diagram illustrating an example of an estimation result in a comparative example. [Figure 9] FIG. 4 is a diagram showing a first example of an estimation result according to the present embodiment. [Figure 10] FIG. 10 is a diagram showing a second example of an estimation result according to the present embodiment. [Figure 11] FIG. 1 shows an example of the correlation between estimated and experimental values ​​of binding energy at a ligand binding site. [Figure 12] FIG. 10 is a diagram illustrating an example of a procedure for generating a learning model by an information processing device. [Figure 13] FIG. 10 is a diagram showing an example of a procedure for estimating intermolecular interaction energy by an information processing device. DETAILED DESCRIPTION OF THE INVENTION

[0010] The present invention will be described below with reference to the drawings showing embodiments thereof. FIG. 1 is a diagram showing an example of the configuration of an information processing device 50. The information processing device 50 can be connected to data servers 100 and 200 via a communication network 1. The data server 100 includes a polymolecular structure DB (database) 110, and the data server 200 includes a single-molecule structure DB 210. The information processing device 50 can acquire information recorded in the polymolecular structure DB 110 and the single-molecule structure DB 210 by accessing the data servers 100 and 200. Note that the information processing device 50 may be configured with a plurality of information processing devices, for example, with functions distributed among them.

[0011] FIG. 2 is a diagram showing an example of information recorded in the multi-molecular system structure DB110. The multi-molecular system structure DB110 can be, for example, a database in which a dataset of solvated protein fragments (SPFs) is recorded. Note that the multi-molecular system structure DB110 is not limited to SPFs, and a dataset of a multi-molecular system structure other than SPFs can be set. The multi-molecular system structure DB110 may be, for example, a database of fragmented nucleic acids. A multi-molecular system structure is configured as a system of multiple molecules and is composed of structures of multiple molecules. The multi-molecular system structure DB110 (SPF) records the molecular formation energy of a multi-molecular system structure between multiple protein fragment (amino acid) molecules and water molecules, including intermolecular interaction energy, in association with the feature quantities of the multi-molecular system structure. The multi-molecular system structure includes approximately 2.7 million structures. A complex is a structure in which two or more molecules are bound together. Typical intermolecular interactions include electrostatic interactions, van der Waals interactions, and hydrogen-bond interactions, but are not limited to these. Other interactions include, for example, hydrogen-bond-like interactions, π-π interactions, coordination bond interactions, and charge-transfer interactions.

[0012] As shown in Figure 2, a protein fragment as an example of a multi-molecular system structure is associated with the atomic numbers and coordinates of the atoms constituting the protein fragment as feature quantities, and the molecular formation energy of the protein fragment. Note that the multi-molecular system structure is not limited to protein fragments, but also includes, for example, multi-molecular systems such as acetic acid and ethanol. The feature quantities of a multi-molecular complex include the atomic numbers and coordinates (atomic arrangement) of the atoms constituting the molecules contained in the complex. Furthermore, the feature quantities of a molecule contained in the complex include the atomic numbers and coordinates of the atoms constituting the molecule. The molecular formation energy is the molecular formation energy based on the results of quantum chemical calculations. The molecular formation energy of a molecular system is the energy required to generate the molecular system from the energies of isolated atoms constituting the molecular system. If the sum of the energies of the isolated atoms is Ea and the energy of the molecular system is Em, the molecular formation energy E can be calculated using the formula E = Em - Ea. Note that the molecular formation energy can also be referred to as atomization energy, bond energy, or dissociation energy.

[0013] As shown in FIG. 2, a certain protein fragment is denoted by the symbol PF0001 for convenience. The atomic numbers of the atoms constituting the protein fragment PF0001 are denoted by the symbols A, B, and C for convenience. The molecular formation energy of the protein fragment PF0001 is denoted by the symbol E(PF0001) for convenience. As a result, the data set contains information in which the protein fragment PF0001, the atomic numbers of atoms A, B, and C, the coordinates of atoms A, B, and C, and the molecular formation energy E(PF0001) of the protein fragment PF0001 are associated with each other. The same applies to other protein fragments. Note that in addition to the molecular formation energy, force, dipole moment, etc. may also be added. The molecular formation energy may be referred to as a molecular property, or the molecular property may be defined by adding force and dipole moment.

[0014] In general machine learning (for example, GNN: Graph Neural Network), a molecular graph is created with the atoms that make up the molecule as nodes and the bonds between the atoms as edges, and information on chemical bonds such as single bonds and double bonds is used as feature quantities for the edges. On the other hand, the dataset of the polymolecular system structure DB110 (SPF), which is one of the embodiments of the present invention, uses the atomic numbers and coordinates of atoms as feature quantities as described above, and does not use edge information based on chemical bonds. In other words, since the information between atoms simply considers only the distance between atoms, it can also be applied to polymolecular systems.

[0015] FIG. 3 is a diagram showing an example of information recorded in the single-molecule structure DB210. The single-molecule structure DB210 can be, for example, a database in which a QM9 dataset is recorded. Note that the single-molecule structure DB210 is not limited to QM9 and may be PubChemQC, etc. A single-molecule structure is a system consisting of one molecule and is composed of the structure of one molecule. The single-molecule structure DB210 (QM9) records, for each molecule (single molecule), the atomic numbers and coordinates (atomic arrangement) of the atoms constituting the molecule as molecular feature quantities, as well as the molecular formation energy of the molecule, in association with each other. The single-molecule structure DB210 includes approximately 130,000 types of molecular structures. The atoms constituting the molecule are, for example, hydrogen atoms, carbon atoms, oxygen atoms, nitrogen atoms, and fluorine atoms. As with the multi-molecular structure shown in Figure 2, the molecular formation energy of a molecule is the energy required to create a molecule from the energy of isolated atoms that make up the molecule. If the sum of the energies of the isolated atoms is Ea and the energy of the molecule is Em, the molecular formation energy E can be calculated using the formula E = Em - Ea. The molecular formation energy is calculated using the results of quantum chemical calculations. All single-molecule structures are the most stable structures.

[0016] As shown in FIG. 3, a certain molecule is represented by the symbol M0001 for convenience. The atomic numbers of the atoms constituting molecule M0001 are represented by the symbols A and B for convenience. Furthermore, the molecular formation energy of molecule M0001 is represented by the symbol E(M0001) for convenience. As a result, the data set is information in which the molecule M0001, the atomic numbers of atoms A and B, the coordinates of atoms A and B, and the molecular formation energy E(M0001) of molecule M0001 are associated with each other. The same applies to other molecules. In addition to the molecular formation energy, the force acting between molecules, the intermolecular dipole moment, etc. may also be added. The molecular formation energy may be referred to as a molecular characteristic, or the force acting between molecules and the intermolecular dipole moment may also be added to the molecular characteristics.

[0017] As shown in FIG. 1, the information processing device 50 includes a control unit 51 that controls the entire device, a communication unit 52, a memory 53, a display unit 54, an operation unit 55, a storage unit 56, and an estimation unit 59.

[0018] The control unit 51 may be configured by incorporating a required number of central processing units (CPUs), micro-processing units (MPUs), graphics processing units (GPUs), etc. The control unit 51 may also be configured by combining digital signal processors (DSPs), field-programmable gate arrays (FPGAs), etc.

[0019] The communication unit 52 includes a communication module and has a function of communicating with the data servers 100 and 200 via the communication network 1 .

[0020] The display unit 54 is configured with a liquid crystal display, an organic EL display, or the like, and provides a UI (user interface) to the user by displaying required information. The display unit 54 is equipped with a microphone and a speaker, and can input and output audio. Note that an external display device may be used instead of the display unit 54.

[0021] The operation unit 55 is configured, for example, by a touch panel, and can operate icons displayed on the display unit 54, move and operate a cursor, input characters, etc. The operation unit 55 may be configured by buttons, switches, etc., or may be configured by a keyboard, mouse, etc. The operation unit 55 provides a UI to the user by accepting user operations. Note that an external terminal device for operation may be provided instead of the operation unit 55.

[0022] The storage unit 56 can be configured with a semiconductor memory, a hard disk, or the like, and stores a computer program (program product) 57, a learning model 58, and required information.

[0023] The computer program 57 can be stored in the storage unit 56 by reading the computer program 57 recorded on a recording medium (for example, an optically readable disk storage medium such as a CD-ROM) M using a recording medium reading unit (not shown). The computer program 57 may also be read from a recording medium such as a storage device (semiconductor memory such as a solid state drive (SSD)) connected according to a standard for connecting to a computer (for example, USB (Universal Serial Bus) or other standard) and stored in the storage unit 56. The computer program 57 may also be downloaded from an external device via the communication unit 52 and stored in the storage unit 56.

[0024] When the feature quantities of a multi-molecular complex are input, the learning model 58 can output the molecular generation energy (also referred to as molecular characteristics) of the complex. Furthermore, when the feature quantities of the molecules contained in the complex are input, the learning model 58 can output the molecular generation energy (also referred to as molecular characteristics) of the molecules. The learning model 58 can use, for example, models such as SchNet, PaiNN, PhysNet, and MPNN (Message Passing Neural Network), but is not limited to these.

[0025] The estimation unit 59 estimates the intermolecular interaction energy of two or more (plural) molecules (molecular system) included in the complex. For example, the interaction energy ΔEij between two molecules can be expressed as ΔEij = Ec - (Ei + Ej). Here, Ec is the molecular formation energy of the complex, and Ei and Ej are the molecular formation energies of molecule i and molecule j included in the complex, respectively.

[0026] Similarly, the interaction energy ΔEijk between three molecules can be expressed as ΔEijk=Ec-(Ei+Ej+Ej)-(ΔEij+ΔEjk+ΔEki). Here, Ec is the molecular formation energy of the complex, and Ei, Ej, and Ek are the molecular formation energies of molecules i, j, and k contained in the complex, respectively. Also, ΔEij, ΔEjk, and ΔEki are the interaction energies between two molecules among the three molecules.

[0027] The estimation unit 59 can estimate the intermolecular interaction energy ΔE of the complex by the formula ΔE=Ec-(Ei+ΔEi), where ΔEi is the intermolecular interaction energy of each molecule constituting the complex.

[0028] The memory 53 can be configured with a semiconductor memory such as a static random access memory (SRAM), a dynamic random access memory (DRAM), or a flash memory. A computer program 57 can be loaded into the memory 53, and the control unit 51 can execute the computer program 57. The control unit 51 can execute processing defined by the computer program 57. In other words, processing by the control unit 51 is also processing by the computer program 57.

[0029] As used herein, a monomolecular system refers to a system consisting of only one molecule. A monomolecular system includes molecules having a known typical molecular skeleton. Examples of monomolecules include, but are not limited to, ammonia molecules, water molecules, and methane molecules. A multimolecular system refers to a system consisting of two or more molecules. A multimolecular system includes a combination of monomolecules. Examples of multimolecules include, but are not limited to, a system of an ammonia molecule and a methane molecule, a system of an ammonia molecule, a water molecule, and a methane molecule, and a system of a dimethylethylene molecule and three water molecules. A complex refers to a complex consisting of a protein and a drug molecule, and includes multiple molecules. Note that a complex may also include a complex consisting of multiple molecules.

[0030] As described above, the information processing device 50 (control unit 51) acquires features of the complex and the molecules contained in the complex, and when the features of the complex and the molecules contained in the complex are input, the acquired features are input into the learning model 58, which outputs the molecular generation energy of the complex and the molecules contained in the complex, to acquire the molecular generation energy of the complex and the molecules contained in the complex output by the learning model 58, and the interaction energy between the molecules contained in the complex can be estimated based on the acquired molecular generation energy (by inputting the acquired molecular generation energy into the estimation unit 59).

[0031] When calculating the intermolecular interaction energy of molecules contained in a complex using quantum chemical calculations, the calculation time can take several tens of hours. In drug discovery, when searching to evaluate drug compounds that interact with target proteins, there are a large number of drug compounds to be evaluated (for example, up to 10 60 ), long calculation times result in increased development costs for pharmaceutical development. However, according to this embodiment, by using the learning model 58, it is not necessary to perform quantum chemical calculations when estimating the intermolecular interaction energy, and therefore calculations of intermolecular interactions can be performed faster than quantum chemical calculations.

[0032] Quantum chemical calculations are a method for analyzing the structure and properties of atoms and molecules from their electronic states, and use the Schrödinger equation, which describes the behavior of electrons. In actual calculations, some kind of approximation is introduced to solve the Schrödinger equation, but molecular properties can be calculated from the type and arrangement of atoms.

[0033] Next, a method for generating (learning) the learning model 58 will be described.

[0034] 4 is a diagram showing a first example of a method for generating a learning model 58. The control unit 51 acquires training data from the polymolecular system structure DB 110. The training data includes learning input data and teacher data. The control unit 51 can acquire evaluation data for evaluating the generated learning model 58 from the polymolecular system structure DB 110. The evaluation data also includes learning input data and teacher data, similar to the training data.

[0035] The learning input data includes feature amounts of a complex or feature amounts of molecules contained in the complex. The feature amounts include atomic numbers and coordinates of atoms constituting the molecule. The complex is a complex made up of multiple molecules.

[0036] The training data includes the molecular formation energy of the complex or the molecular formation energy of the molecules contained in the complex.

[0037] The control unit 51 inputs the feature amounts of the complex (learning input data) into the learning model 58, and adjusts the parameters of the learning model 58 so that the error between the molecular formation energy of the complex output by the learning model 58 and the molecular formation energy of the complex as training data is small (so that it is within an acceptable range), thereby generating (learning) the learning model 58. In this way, when the feature amounts of the complex are input, the learning model 58 can output the molecular formation energy of the complex.

[0038] Furthermore, the control unit 51 inputs the feature quantities (learning input data) of the molecules contained in the complex into the learning model 58, and adjusts the parameters of the learning model 58 so that the error between the molecular formation energy of the molecule output by the learning model 58 and the molecular formation energy of the molecule as training data is small (so that it is within an acceptable range), thereby generating (learning) the learning model 58. In this way, when the feature quantities of the molecules contained in the complex are input, the learning model 58 can output the molecular formation energy of the molecule.

[0039] As described above, the control unit 51 acquires training data including the feature amounts of a complex and the molecules contained in the complex, and the molecular formation energies of the complex and the molecules contained in the complex, and can generate the learning model 58 based on the training data so that when the feature amounts of the complex are input, the learning model 58 outputs the molecular formation energies of the complex. Furthermore, the control unit 51 can generate the learning model 58 based on the training data so that when the feature amounts of the molecules contained in the complex are input, the learning model 58 outputs the molecular formation energies of the molecules contained in the complex.

[0040] The training data may also include the feature quantities and molecular formation energies of the complex and the molecules contained in the complex.

[0041] 5 is a diagram showing a second example of a method for generating a learning model 58. The control unit 51 acquires training data from the single-molecule structure DB 210. The training data includes learning input data and teacher data. The control unit 51 can also acquire evaluation data for evaluating the generated learning model 58 from the single-molecule structure DB 210. Like the training data, the evaluation data also includes learning input data and teacher data.

[0042] The learning input data includes feature quantities of molecules in a single molecule system, including atomic numbers and coordinates of atoms that constitute the molecules.

[0043] The training data includes the molecular formation energy of molecules in a monomolecular system.

[0044] The control unit 51 inputs the molecular features (learning input data) of the single-molecule system into the learning model 58, and adjusts the parameters of the learning model 58 so that the error between the molecular generation energy output by the learning model 58 and the molecular generation energy as training data is small (within an acceptable range), thereby generating (learning) the learning model 58.

[0045] As described above, the control unit 51 acquires training data including molecular features and molecular generation energy of a single molecule system, and based on the training data, can generate a learning model 58 so that when the features of a single molecule are input, the learning model 58 outputs the molecular generation energy of the single molecule.

[0046] The training data may also include molecular features and molecular formation energies of the single-molecule structure.

[0047] In the above example, the information processing device 50 is configured to generate the learning model 58, but this is not limiting. For example, the learning model 58 may be generated by a learning processing device separate from the information processing device 50, and the generated learning model 58 may be stored in the storage unit 56 of the information processing device 50.

[0048] Next, a method for estimating the interaction energy between molecules (also referred to as intermolecular interaction energy) will be described.

[0049] FIG. 6 is a diagram illustrating a first example of a method for estimating intermolecular interaction energy using the information processing device 50. In the example illustrated in FIG. 6, water molecules and ethanol molecules are used as molecules contained in a complex, and a complex of water molecules and ethanol molecules is used as a complex. First, the control unit 51 inputs the feature quantities of the water molecules into the learning model 58 to obtain the molecular formation energy E1 of the water molecules output by the learning model 58. The feature quantities of the water molecules are the atomic numbers and coordinates of the hydrogen and oxygen atoms. Next, the control unit 51 inputs the feature quantities of the ethanol molecules into the learning model 58 to obtain the molecular formation energy E2 of the ethanol molecules output by the learning model 58. The feature quantities of the ethanol molecules are the atomic numbers and coordinates of the carbon, hydrogen, and oxygen atoms. Next, the control unit 51 inputs the feature quantities of the complex of water molecules and ethanol molecules into the learning model 58 to obtain the molecular formation energy E3 of the complex output by the learning model 58. The feature quantities of the complex are the atomic numbers and coordinates of the carbon, hydrogen, and oxygen atoms. Note that the order in which the feature quantities are input into the learning model 58 is not particularly specified.

[0050] The control unit 51 inputs the acquired molecule formation energies E1, E2, and E3 to the estimation unit 59. The estimation unit 59 estimates the interaction energy ΔE between the molecules contained in the complex (i.e., between the water molecules and the ethanol molecules) based on the input molecule formation energies E1, E2, and E3. The interaction energy ΔE can be calculated using the formula ΔE=E3-E1-E2.

[0051] FIG. 7 is a diagram illustrating a second example of a method for estimating intermolecular interaction energy using the information processing device 50. In the example illustrated in FIG. 7, the molecules contained in the complex are a protein as a target and a ligand molecule that binds to the protein, and the complex is an example of a complex consisting of a protein and a ligand molecule. The ligand molecule binds to the protein to form a complex. First, the control unit 51 inputs the feature values ​​of the protein into the learning model 58 to obtain the molecular formation energy Ep of the protein alone, which is output by the learning model 58. The feature values ​​of the protein are the atomic number and coordinates of each atom that constitutes the protein. Next, the control unit 51 inputs the feature values ​​of the ligand molecule into the learning model 58 to obtain the molecular formation energy El of the ligand molecule, which is output by the learning model 58. The feature values ​​of the ligand molecule are the atomic number and coordinates of each atom that constitutes the ligand molecule. Next, the control unit 51 inputs the feature values ​​of the complex between the protein and the ligand molecule into the learning model 58 to obtain the molecular formation energy Ec of the complex, which is output by the learning model 58. The feature values ​​of the complex are the atomic number and coordinates of each atom that constitutes the protein and the ligand molecule. Note that the order in which the feature values ​​are input into the learning model 58 is not particularly specified.

[0052] The control unit 51 inputs the acquired molecule formation energies Ep, El, and Ec to the estimation unit 59. Based on the input molecule formation energies Ep, El, and Ec, the estimation unit 59 estimates the interaction energy ΔE between the molecules contained in the complex (i.e., between the protein and the ligand molecule). The interaction energy ΔE can be calculated by the formula ΔE=Ec-Ep-El.

[0053] As described above, the control unit 51 acquires features of a complex of a protein and a ligand molecule that binds to the protein, as well as the protein and the ligand molecule, inputs the acquired features into the learning model 58, and acquires the molecular formation energies (Ec, Ep, and El) of the complex, the protein, and the ligand molecule that are output by the learning model 58. The control unit 51 inputs the acquired molecular formation energies into the estimation unit 59, thereby estimating the intermolecular interaction energy ΔE between the protein and the ligand molecule contained in the complex. In other words, the complex includes a protein and a ligand molecule that binds to the protein, and the control unit 51 can estimate the interaction energy between the protein and the ligand molecule. The features also include the coordinates and atomic numbers of the atoms that constitute the molecules contained in the complex.

[0054] As a result, according to this embodiment, there is no need to perform quantum chemical calculations at the stage of estimating intermolecular interaction energy, and the intermolecular interaction between a target protein and a ligand molecule (drug compound) that interacts with it can be estimated more quickly than when performing quantum chemical calculations. This makes it possible to speed up the evaluation of many drug compounds in drug development in drug discovery, contributing to shortening the development period and reducing development costs for drug development.

[0055] Next, evaluation of the estimation result by the information processing device 50 of this embodiment will be described.

[0056] FIG. 8 is a diagram showing an example of the estimation results of a comparative example. In FIG. 8, the vertical axis represents intermolecular interaction [kcal / mol], and the horizontal axis represents intermolecular distance [Å]. FIG. 8 illustrates the intermolecular interaction energy between water molecules and ethanol molecules. In the figure, the X mark indicates the intermolecular interaction energy based on the quantum chemistry calculation results, and the △ mark indicates the intermolecular interaction energy obtained by the learning model 58 trained using only single-molecule system data. In this case, the intermolecular interaction energy based on the quantum chemistry calculation results can be considered to be the correct value.

[0057] As shown in FIG. 8, when the intermolecular distance is longer than 3.0 Å, the intermolecular interaction energy calculated by the learning model 58 and the quantum chemical calculation result are nearly identical, and it can be said that the prediction (estimation) by the learning model 58 is successful. On the other hand, when the intermolecular distance is shorter than 3.0 Å, the intermolecular interaction energy calculated by the learning model 58 and the quantum chemical calculation result do not match, and the shorter the intermolecular distance, the greater the difference between the two values. In other words, it can be seen that the intermolecular interaction energy calculated by the learning model 58 trained only with single-molecule system data becomes smaller, and the intermolecular bond becomes more stable. In other words, it is inferred that the learning model 58 has not been able to learn the repulsive force between water molecules and ethanol molecules. In this embodiment, the learning model 58 is trained using multi-molecule system data so that it can learn the intermolecular repulsive force.

[0058] Furthermore, by using a learning model 58 trained only on single-molecule data, quantum chemical calculations can be substituted for single small molecules, and molecular properties such as the forces acting between molecules and the intermolecular dipole moments can be estimated with high accuracy.

[0059] FIG. 9 is a diagram showing a first example of the estimation results of this embodiment. In FIG. 9, the vertical axis represents intermolecular interaction [kcal / mol], and the horizontal axis represents intermolecular distance [Å]. FIG. 9 illustrates the intermolecular interaction energy between water molecules and ethanol molecules. In the diagram, an X symbol indicates the intermolecular interaction energy based on the quantum chemistry calculation results, a △ symbol indicates the intermolecular interaction energy obtained by the learning model 58 trained only with single-molecule system data, and an O symbol indicates the intermolecular interaction energy obtained by the learning model 58 trained only with multi-molecule system data. In this case, as in the case of FIG. 8, the intermolecular interaction energy based on the quantum chemistry calculation results can be regarded as the correct value. Note that the intermolecular interaction energy based on the quantum chemistry calculation results and the intermolecular interaction energy obtained by the learning model 58 trained only with single-molecule system data are the same as in the example of FIG. 8.

[0060] 9, regardless of the length of the intermolecular distance, the intermolecular interaction energy obtained by the learning model 58 trained only with multi-molecule system data is almost identical to the quantum chemical calculation results, and it can be said that the prediction (estimation) by the learning model 58 is successful. That is, the MAE (Mean Absolute Error) when learning only with single-molecule system data is 8.0995, while the MAE when learning only with multi-molecule system data is 0.7330, and the estimation accuracy of intermolecular interactions is higher when learning only with multi-molecule system data than when learning only with single-molecule system data.

[0061] By using the learning model 58 that has been trained using only multi-molecular system data, it becomes possible to estimate the interaction energy of a multi-molecular system with high accuracy.

[0062] Fig. 10 is a diagram showing a second example of the estimation results of this embodiment. In Fig. 10, the intermolecular interaction energy obtained by the learning model 58 trained with both data sets of multi-molecule system data and single-molecule system data is indicated by a + sign in the example of Fig. 9. As shown in Fig. 10, it can be seen that the intermolecular interaction energy obtained by the learning model 58 trained with both data sets is closer to the quantum chemical calculation result in the extremely short distance range of intermolecular distances around 2.0 Å than the intermolecular interaction energy obtained by the learning model 58 trained only with multi-molecule system data.

[0063] As described above, by using a learning model 58 trained only with multi-molecule system data, or a learning model 58 trained with both multi-molecule system data and single-molecule system data, it is possible to accurately estimate the intermolecular interaction energy.

[0064] Figure 11 shows an example of the correlation between estimated and experimental values ​​of binding energy at a ligand binding site. In Figure 11, the vertical axis represents experimental binding energy, and the horizontal axis represents estimated (calculated) binding energy. In the figure, symbols L12, L13, L14, L20, LG2, LG3, LG5, LG6, LG8, and LG9 represent the types of ligands. Existing QM methods use quantum chemical calculations, while QM-AI uses a learning model58 and does not use chemical bond information.

[0065] 11, the correlation coefficient between the experimental value and the calculated value in the cases of Comparative Examples 1 and 2 (existing QM methods) was 0.77 for Comparative Example 1 and 0.79 for Comparative Example 2. On the other hand, the correlation coefficient between the experimental value and the estimated value (calculated value) when learning model 58 trained only with multi-molecular system data was used was 0.74, and the correlation coefficient between the experimental value and the estimated value (calculated value) when learning model 58 trained with both data sets of multi-molecular system data and single-molecular system data was used was 0.85.

[0066] As described above, in this embodiment, it is possible to use the learning model 58 trained only with multi-molecule system data, or the learning model 58 trained with both data sets of multi-molecule system data and single-molecule system data. However, it is found that the use of the learning model 58 trained with both data sets of multi-molecule system data and single-molecule system data further improves the correlation between the experimental value and the estimated value (calculated value).

[0067] 12 is a diagram showing an example of the procedure for generating a learning model 58 by the information processing device 50. The control unit 51 acquires training data including the feature quantities of a complex and the molecules contained in the complex, and the molecular formation energy, from the multi-molecular structure DB 110 (SPF) (S11). Based on the acquired training data, the control unit 51 generates a learning model 58 so that, when the feature quantities of a complex are input, the molecular formation energy of the complex is output (S12).

[0068] When the control unit 51 receives the feature amount of a molecule contained in a complex based on the acquired training data, it generates a learning model 58 so as to output the molecular formation energy of the molecule (S13).

[0069] The control unit 51 acquires training data including the complex, the molecules contained in the complex, the features of the single molecules, and the molecular formation energies from the multi-molecular structure DB 110 (SPF) and the single-molecular structure DB 210 (QM9) (S14). Based on the acquired training data, the control unit 51 generates a learning model 58 so that, when the features of the complex, the molecules contained in the complex, and the single molecules are input, the control unit 51 outputs the molecular formation energies of the complex, the molecules contained in the complex, and the single molecules (S15). The control unit 51 stores the generated learning model 58 in the memory unit 56 (S16) and ends the process.

[0070] Although the processing of steps S14 and S15 is not essential, the accuracy of estimation of the intermolecular interaction energy by the learning model 58 can be further improved by using the data set of the single-molecule structure DB210 (QM9) as training data in addition to the data set of the multi-molecular structure DB110 (SPF).

[0071] 13 is a diagram showing an example of the procedure for estimating intermolecular interaction energy by the information processing device 50. The control unit 51 acquires features of the complex (S21), and acquires features of the molecules contained in the complex (S22). The control unit 51 inputs the features of the complex into the learning model 58, and acquires the molecular formation energy of the complex output by the learning model 58 (S23).

[0072] The control unit 51 inputs the feature amounts of the molecules contained in the complex into the learning model, and obtains the molecular formation energy of the molecules output by the learning model 58 (S24). The control unit 51 estimates the intermolecular interaction energy of the complex based on the molecular formation energy of the complex and the molecular formation energy of the molecules contained in the complex (S25), and ends the process.

[0073] (Supplementary Note 1) The computer program causes a computer to execute a process of acquiring features of a complex and molecules contained in the complex, inputting the acquired features into a learning model that outputs molecular formation energies of the complex and the molecules contained in the complex when the features of the complex and the molecules contained in the complex are input, acquiring the molecular formation energies of the complex and the molecules contained in the complex output by the learning model, and estimating the interaction energy between the molecules contained in the complex based on the acquired molecular formation energies.

[0074] (Appendix 2) The computer program in Appendix 1 causes a computer to execute a process in which the complex includes a protein and a ligand molecule that binds to the protein, and the process estimates the interaction energy between the protein and the ligand molecule.

[0075] (Supplementary Note 3) In the computer program according to Supplementary Note 1 or Supplementary Note 2, the feature amount includes coordinates and atomic numbers of atoms that constitute molecules contained in the complex.

[0076] (Supplementary Note 4) The information processing device includes a control unit, which acquires features of a complex and molecules contained in the complex, and when the control unit inputs the features of the complex and molecules contained in the complex, inputs the acquired features into a learning model that outputs molecular generation energies of the complex and molecules contained in the complex, acquires the molecular generation energies of the complex and molecules contained in the complex output by the learning model, and estimates the interaction energy between the molecules contained in the complex based on the acquired molecular generation energies.

[0077] (Appendix 5) An information processing method acquires features of a complex and molecules contained in the complex, and when the features of the complex and the molecules contained in the complex are input, the acquired features are input into a learning model that outputs molecular formation energies of the complex and the molecules contained in the complex, and the molecular formation energies of the complex and the molecules contained in the complex output by the learning model are acquired, and an interaction energy between the molecules contained in the complex is estimated based on the acquired molecular formation energies.

[0078] (Appendix 6) A learning model generation method acquires training data including features of a complex and molecules contained in the complex, and molecular formation energies of the complex and the molecules contained in the complex, and generates a learning model based on the training data so that when the features of the complex are input, the learning model outputs the molecular formation energy of the complex, and when the features of the molecules contained in the complex are input, the learning model is generated based on the training data so that the learning model outputs the molecular formation energy of the molecules contained in the complex.

[0079] (Supplementary Note 7) In the learning model generation method according to Supplementary Note 6, the training data includes feature quantities and molecule formation energies of the molecules and complexes of the multi-molecular system structure.

[0080] (Supplementary Note 8) In the learning model generation method according to Supplementary Note 7, the training data further includes feature quantities and molecule formation energies of molecules with single molecular structures.

[0081] (Supplementary Note 9) In the learning model generation method according to any one of Supplementary Note 6 to Supplementary Note 8, the feature amount includes coordinates and atomic numbers of atoms constituting molecules contained in the complex.

[0082] The matters described in each embodiment can be combined with each other. In addition, the independent claims and dependent claims described in the claims can be combined with each other in any combination, regardless of the reference format. Furthermore, although the claims use a format in which a claim references two or more other claims (multiple claim format), this is not limited to this format. A multiple claim (multi-multi claim) that references at least one other multiple claim may also be used. [Explanation of symbols]

[0083] 1. Communication Network 50 Information processing equipment 51 Control section 52 Communications Department 53 Memory 54 Display section 55 Operation section 56 Memory section 57 Computer Programs 58 Learning Model 59 Estimation part 100 Data Servers 110 Multimolecular system structure DB 200 Data Server 210 Single-Molecular Structure DB

Claims

1. Acquire features of the complex and the molecules contained in the complex; When feature amounts of a complex and a molecule contained in the complex are input, the acquired feature amounts are input into a learning model that outputs molecular formation energies of the complex and the molecules contained in the complex, and the molecular formation energies of the complex and the molecules contained in the complex output by the learning model are acquired; Estimating the interaction energy between the molecules contained in the complex based on the obtained molecular formation energy. A computer program that causes a computer to perform a process.

2. The complex is a protein and a ligand molecule that binds to the protein, estimating the interaction energy between the protein and the ligand molecule; 2. The computer program according to claim 1, which causes a computer to execute a process.

3. The feature amount is The complex includes the coordinates and atomic numbers of the atoms constituting the molecule.

3. A computer program according to claim 1 or claim 2.

4. A control unit is provided, The control unit Acquire features of the complex and the molecules contained in the complex; When feature amounts of a complex and a molecule contained in the complex are input, the acquired feature amounts are input into a learning model that outputs molecular formation energies of the complex and the molecules contained in the complex, and the molecular formation energies of the complex and the molecules contained in the complex output by the learning model are acquired; Estimating the interaction energy between the molecules contained in the complex based on the obtained molecular formation energy. Information processing device.

5. Acquire features of the complex and the molecules contained in the complex; When feature amounts of a complex and a molecule contained in the complex are input, the acquired feature amounts are input into a learning model that outputs molecular formation energies of the complex and the molecules contained in the complex, and the molecular formation energies of the complex and the molecules contained in the complex output by the learning model are acquired; Estimating the interaction energy between the molecules contained in the complex based on the obtained molecular formation energy. Information processing methods.

6. Acquire training data including feature amounts of a complex and molecules contained in the complex, and molecular formation energies of the complex and molecules contained in the complex; generating a learning model based on the training data so as to output the molecular formation energy of the complex when a feature amount of the complex is input; generating the learning model based on the training data so as to output molecular formation energies of the molecules included in the complex when feature amounts of the molecules included in the complex are input; Learning model generation method.

7. The training data is The molecular formation energy and the characteristic quantities of each molecule and complex of the multi-molecular system structure are included. The learning model generation method according to claim 6 .

8. The training data is Furthermore, the molecular characteristics of the single molecular structure and the molecular formation energy are included. The learning model generation method according to claim 7 .

9. The feature amount is The complex includes the coordinates and atomic numbers of the atoms constituting the molecule. The learning model generation method according to any one of claims 6 to 8.

Images