Learning device, prediction device, learning method, learning program, prediction method, and prediction program

Abstract

To improve prediction accuracy when predicting an intensity index related to the intensity of a mass spectrum of a compound.SOLUTION: A learning device according to the present invention includes a learning unit that generates a learned model by performing learning using a learning database in which a physical index related to an atom constituting a compound calculated based on a machine learning potential is associated with an intensity index related to an intensity of a mass spectrum of the compound.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a learning device, a prediction device, a learning method, a learning program, a prediction method, and a prediction program. [Background technology]

[0002] When identifying compounds in a sample using mass spectrometry, the mass spectrum obtained by analysis using a mass spectrometer or other device is generally used. Typically, the molecular structure of the unknown compound is estimated by comparing the mass spectrum obtained by mass spectrometry of a sample containing the unknown compound to be identified with standard mass spectra of various compounds stored in a mass spectral library.

[0003] A typical known compound database contains more information about compounds, such as their structures, than a known mass spectral library. However, a known compound database contains a vast number of compounds whose mass spectra are not included in a known mass spectral library. A representative known mass spectral library, for example, is NIST20, a database provided by the National Institute of Standards and Technology (NIST), which contains EI spectra for over 300,000 compounds and MS / MS spectra for over 30,000 compounds. Meanwhile, a representative known compound database, for example, PubChem, contains the structures of over 100 million compounds. Therefore, when identifying unknown compounds using mass spectra, it is often difficult to find a matching molecular structure from a mass spectral library, making it impossible to identify the unknown compound.

[0004] One solution to obtaining mass spectra for compounds for which matching molecular structures cannot be found in mass spectral libraries is to expand the database of compounds not included in mass spectral libraries. However, it is not practical to synthesize the vast number of compounds not included in mass spectral libraries and obtain their mass spectra. Furthermore, when identifying the molecular structure of an unknown compound from its mass spectrum, if the structure is simple, it is possible to identify the molecular structure of the unknown compound by matching the mass spectrum of the unknown compound with fragments of candidate molecules. However, such analysis is extremely difficult for non-experts, and even experts have difficulty identifying complex structures. Therefore, as a solution to obtaining mass spectra for compounds for which matching molecular structures cannot be found in mass spectral libraries, the development of prediction methods that can predict the mass spectrum of a compound from its molecular structure has been investigated.

[0005] As a prediction method for predicting the mass spectrum of a compound from its molecular structure, for example, a method for predicting MS / MS spectra using a graph convolutional neural network (GCNN) or a multilayer perceptron (MLP) trained by inputting molecular structure information has been disclosed (see, for example, Non-Patent Documents 1 and 2). [Prior art documents] [Non-patent literature]

[0006] [Non-Patent Document 1] J. Park et al., Scientific Reports, 14, 1400 (2024) [Non-patent document 2] A. Young et al., "MassFormer: Tandem Mass Spectrum Prediction for Small Molecules using Graph Transformers", arXiv:2111.04824v3, 1 May, 2023 Summary of the Invention [Problem to be solved by the invention]

[0007] However, the methods of Non-Patent Documents 1 and 2 have a problem in that there is a large error between the predicted value of the mass spectrum intensity of the predicted compound and the experimental value indicating the mass spectrum intensity of the compound obtained by experiment, and the prediction accuracy of indices related to the mass spectrum intensity, such as the mass spectrum intensity of the predicted compound, is low.

[0008] An object of the present invention is to improve the accuracy of prediction when predicting intensity indices relating to the intensity of the mass spectrum of a compound. [Means for solving the problem]

[0009] One aspect of the present invention is This is a learning device that includes a learning unit that generates a trained model by learning using a learning database in which physical indices related to the atoms that constitute a compound, calculated based on machine learning potential, are associated with intensity indices related to the intensity of the mass spectrum of the compound.

[0010] Another aspect of the present invention is The prediction device includes a prediction unit that predicts the intensity index related to the intensity of the mass spectrum of a compound to be predicted by inputting the physical index of the compound to be predicted and structural features of the compound to be predicted into a trained model that has been trained using a training database in which physical indexes related to the atoms that constitute the compound, calculated based on machine learning potential, are associated with intensity indexes related to the intensity of the mass spectrum of the compound.

[0011] Another aspect of the present invention is The computer This is a learning method that performs a learning process to generate a trained model by learning using a learning database in which physical indicators related to the atoms that constitute a compound, calculated based on the machine learning potential, are associated with intensity indicators related to the intensity of the mass spectrum of the compound.

[0012] Another aspect of the present invention is On the computer, This is a learning program that executes a learning process to generate a trained model by learning using a learning database in which physical indicators related to the atoms that constitute a compound, calculated based on the machine learning potential, are associated with intensity indicators related to the intensity of the mass spectrum of the compound.

[0013] Another aspect of the present invention is The computer This prediction method executes a prediction step of predicting the intensity index related to the intensity of the mass spectrum of a compound to be predicted by inputting the physical index of the compound to be predicted and structural features of the compound to be predicted into a trained model trained using a training database in which physical indexes related to the atoms constituting the compound, calculated based on machine learning potential, and intensity indexes related to the intensity of the mass spectrum of the compound are associated with each other.

[0014] Another aspect of the present invention is On the computer, This is a prediction program that executes a prediction step of predicting the intensity index related to the intensity of the mass spectrum of a compound to be predicted by inputting the physical index of the compound to be predicted and structural features of the compound to a trained model that has been trained using a training database in which physical indexes related to the atoms that constitute the compound, calculated based on machine learning potential, and intensity indexes related to the intensity of the mass spectrum of the compound are associated with each other. [Effects of the Invention]

[0015] According to the present invention, it is possible to improve the prediction accuracy when predicting an intensity index relating to the intensity of a mass spectrum of a compound. [Brief explanation of the drawings]

[0016] [Figure 1]FIG. 1 is an explanatory diagram showing an overview of the flow for predicting the mass spectrum intensity of a compound using physical indices and structural features obtained from the molecular structure of the compound in a trained model. [Figure 2] 1 is a block diagram showing a schematic configuration of a learning device according to a first embodiment of the present invention. [Figure 3] FIG. 1 is a diagram illustrating an example of a molecular structure. [Figure 4] FIG. 1 is a diagram illustrating molecular bond energy. [Figure 5] FIG. 1 is a diagram illustrating an example of charges of atoms constituting a compound. [Figure 6] FIG. 2 is a diagram illustrating the molecular weights of the first fragment and the second fragment. [Figure 7] FIG. 2 is a block diagram showing a schematic configuration of a graph database creation unit. [Figure 8] FIG. 4 is a diagram showing an example of a graph structure obtained by converting the molecular structure shown in FIG. 3. [Figure 9] FIG. 1 is a diagram showing an example of the mass spectrum intensity of a compound. [Figure 10] 1 is a block diagram showing a schematic configuration of a prediction device according to a first embodiment of the present invention. [Figure 11] FIG. 2 is a block diagram showing the hardware configuration of a learning device and a prediction device. [Figure 12] 3 is a flowchart illustrating a learning method according to the first embodiment of the present invention. [Figure 13] 10 is a flowchart illustrating a process for creating a graph database. [Figure 14] 1 is a flowchart showing a prediction method according to a first embodiment of the present invention. [Figure 15] FIG. 10 is a block diagram showing a schematic configuration of a prediction device according to a second embodiment of the present invention. [Figure 16] 10 is a flowchart showing a prediction method according to a second embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0017] Hereinafter, embodiments of the present invention will be described in detail. In this specification, unless otherwise specified, the symbol "to" indicating a numerical range means that the numerical values ​​before and after it are included as the lower and upper limits. Furthermore, when a unit is specified for only the upper limit of a numerical range expressed by "to," it means that the lower limit is also expressed in the same unit.

[0018] The learning device of this embodiment includes a learning unit that generates a trained model by learning using a learning database in which physical indices related to the atoms constituting a compound calculated based on the machine learning potential are associated with intensity indices related to the intensity of the mass spectrum of the compound.

[0019] The learning device according to this embodiment obtains highly accurate physical indices by calculating physical indices related to atoms constituting a compound using a machine learning potential. The learning device according to this embodiment performs learning in a learning unit using a learning database that has, as explanatory variables, highly accurate physical indices calculated using the machine learning potential and intensity indices related to the intensities of the mass spectra of compounds as objective variables, thereby generating a trained model with improved prediction accuracy when predicting intensity indices related to the intensities of the mass spectra of various compounds.

[0020] The prediction device according to this embodiment includes: The prediction unit predicts the intensity index related to the intensity of the mass spectrum of the compound to be predicted by inputting the physical index of the compound to be predicted and the structural feature amount of the compound to be predicted into a trained model trained using a training database in which physical indexes related to the atoms constituting the compound, calculated based on the machine learning potential, are associated with intensity indexes related to the intensity of the mass spectrum of the compound.

[0021] The prediction device according to this embodiment calculates physical indices related to atoms constituting a compound to be predicted using a machine learning potential to obtain highly accurate physical indices. The prediction device according to this embodiment also uses a trained model trained using a training database having, as explanatory variables, highly accurate physical indices calculated using the machine learning potential and intensity indices related to the mass spectrum intensity of the compound as a response variable. The prediction device according to this embodiment can improve the prediction accuracy when predicting intensity indices related to the mass spectrum intensity of various compounds to be predicted by inputting the physical indices of the compound to be predicted into a trained model generated in advance by the prediction unit.

[0022] In the learning method of this embodiment, a computer performs a learning process to generate a trained model by learning using a learning database in which physical indicators related to the atoms constituting a compound calculated based on the machine learning potential are associated with intensity indicators related to the intensity of the mass spectrum of the compound.

[0023] Similar to the learning device according to the present embodiment, the learning method according to the present embodiment uses a learning database having, as explanatory variables, highly accurate physical indices calculated using the machine learning potential and, as objective variables, intensity indices related to the intensities of the mass spectra of compounds, for learning in the learning process. As a result, the learning method according to the present embodiment can generate a trained model with improved prediction accuracy when predicting intensity indices related to the intensities of the mass spectra of various compounds.

[0024] The learning program of this embodiment causes a computer to execute a learning process to generate a trained model by learning using a learning database in which physical indicators related to the atoms constituting a compound calculated based on the machine learning potential are associated with intensity indicators related to the intensity of the mass spectrum of the compound.

[0025] Similar to the learning device according to the present embodiment, the learning program according to the present embodiment executes learning in a learning process using a learning database having, as explanatory variables, highly accurate physical indices calculated using the machine learning potential and, as objective variables, intensity indices related to the intensities of the mass spectra of compounds. As a result, the learning program according to the present embodiment can generate trained models that can improve the prediction accuracy when predicting intensity indices related to the intensities of the mass spectra of various compounds.

[0026] In the prediction method according to this embodiment, a computer inputs physical indices of a compound to be predicted into a trained model trained using a training database in which physical indices of atoms constituting the compound, calculated based on a machine learning potential, correspond to intensity indices related to the intensity of the mass spectrum of the compound, thereby executing a prediction step in which the intensity indices related to the intensity of the mass spectrum of the compound are predicted.

[0027] Similar to the prediction device according to the present embodiment, the prediction method according to the present embodiment calculates physical indices related to the atoms constituting the compound to be predicted using a machine learning potential to obtain highly accurate physical indices. Furthermore, the prediction method according to the present embodiment uses a trained model trained using a training database having, as explanatory variables, highly accurate physical indices calculated using the machine learning potential and intensity indices related to the intensity of the mass spectrum of the compound as the objective variable. In the prediction step, the prediction method according to the present embodiment inputs the physical indices of the compound to be predicted into a trained model generated in advance, thereby improving the prediction accuracy when predicting intensity indices related to the intensity of the mass spectrum of various compounds to be predicted.

[0028] The prediction program of this embodiment causes a computer to execute a prediction process in which the intensity index related to the intensity of the mass spectrum of the compound to be predicted is predicted by inputting the physical index of the compound to be predicted into a trained model trained using a training database in which the physical index related to the atoms constituting the compound, calculated based on the machine learning potential, is associated with the intensity index related to the intensity of the mass spectrum of the compound.

[0029] Similar to the prediction device according to the present embodiment, the prediction program according to the present embodiment calculates physical indices related to the atoms constituting the compound to be predicted using a machine learning potential to obtain highly accurate physical indices. Furthermore, the prediction program according to the present embodiment uses a trained model trained using a training database having, as explanatory variables, highly accurate physical indices calculated using the machine learning potential and intensity indices related to the mass spectrum intensity of the compound as the objective variable. Thus, in the prediction step, the prediction program according to the present embodiment can accurately predict intensity indices related to the mass spectrum intensity of various compounds to be predicted by inputting the physical indices of the compound to be predicted into a trained model generated in advance.

[0030] [First embodiment] <Learning device> The learning device according to the present embodiment generates a trained model by learning using a training database in which explanatory variables are associated with objective variables, and the explanatory variables are physical indices of atoms constituting a compound calculated based on a machine learning potential, and the objective variables are intensity indices of the mass spectrum of the compound.

[0031] Figure 1 is an explanatory diagram outlining the flow for predicting the mass spectrum intensity of a compound using physical indices and structural features obtained from the molecular structure of the compound using a trained model. As shown in Figure 1, the molecular structure of a compound is represented as a graph structure (see Figure 1(a)). The vertices (nodes) of the graph structure correspond to each atom of the compound and are assigned features that represent multiple atomic properties, such as atomic number, ionic radius, and charge. The edges of the graph structure correspond to chemical bonds between atoms and are assigned features that include information such as interatomic distance, bond energy, and mass of fragments when the bond is broken. The features of the nodes and edges of the graph structure are represented as vectors in matrix format. In addition, the molecular structure of a compound is represented as a graph structure and as a structural feature vector that represents structural features (e.g., molecular descriptors) (see Figure 1(b)). The node feature vectors and edge feature vectors obtained from the graph structure, along with the structural feature vector, are converted into a feature vector that integrates a large amount of compound information related to mass spectra by processes such as addition, pooling, and combination (see Figure 1(c)). This integrated feature vector is then subjected to nonlinear transformation using a multilayer perceptron (MLP) or similar to convert it into a vector that represents the entire compound in an appropriate matrix format (see Figure 1(d)). The vector that represents the entire compound in an appropriate matrix format is then associated with the intensity of the compound's mass spectrum, creating an intensity index related to the intensity of the compound's mass spectrum (see Figure 1(e)). This results in a training database that associates the physical indices of the atoms that make up the compound and the structural features of the compound with the intensity index related to the intensity of the compound's mass spectrum.

[0032] In the learning device according to this embodiment, the feature of a node, which is used as a physical index for the atoms constituting a compound, is the atomic charge calculated using the machine learning potential, and the feature of an edge, which is used as a feature, is the binding energy calculated using the machine learning potential (see FIG. 1(a)). By using a highly accurate machine learning potential, the feature of the node and the edge can be obtained with high accuracy. A trained model generated by training using a training database in which the physical indexes for the atoms constituting a compound and the structural feature of the compound are associated with the intensity index for the intensity of the mass spectrum of the compound can improve the prediction accuracy when predicting the intensity index for the intensity of the mass spectrum of the compound.

[0033] Fig. 2 is a block diagram showing the schematic configuration of a learning device according to this embodiment. As shown in Fig. 2, the learning device 1 includes a first acquisition unit 11, a graph database creation unit 12, a structure conversion unit 13, a structural feature database creation unit 14, a second acquisition unit 15, a learning unit 16, an output unit 17, a graph database D11, a structural feature database D12, and a learning database D13. The learning device 1 generates a trained model M1 that predicts the mass spectrum of a compound.

[0034] The learning device 1 does not need to have all of the above components, and may be configured to have only the components required depending on the application, etc. For example, it may not be equipped with the structural feature database creation unit 14.

[0035] The first acquisition unit 11 acquires a file relating to the molecular structure of a compound from the molecular information database 110. The molecular structure of a compound may be a molecular structure of a compound that the user has created.

[0036] The file relating to the molecular structure of a compound is a molecular structure file that compiles the molecular structures of compounds corresponding to the database of training mass spectra of the trained model M1 generated by the training unit 16. The file relating to the molecular structure records, for example, the molecular structure of a compound as shown in FIG.

[0037] Files related to molecular structures include mol format files.

[0038] The molecular information database 110 includes a file relating to molecular structure and a file relating to mass spectrum, and the molecular information database 110 may be, for example, an existing database such as NIST 20.

[0039] The graph database creation unit 12 creates a graph database D11 having physical indices related to atoms constituting a compound. The graph database creation unit 12 calculates the physical indices using a machine learning potential based on a file related to a molecular structure acquired by the first acquisition unit 11. The graph database D11 created by the graph database creation unit 12 is used to input into a learning database D13 used when machine learning is performed by the learning unit 16.

[0040] The machine learning potential is an interatomic potential using a machine learning method that outputs energy from information about atomic structures. Examples of machine learning potentials include neural network potential (NNP), Gaussian approximation potential (GAP), spectral neighbor analysis potential (SNAP), and moment tensor potential (MTP). Among these, NNP is preferred as the machine learning potential because of the high flexibility of neural networks. NNP can be an atomic simulator that learns the relationship between atomic coordinates and energy using quantum chemical calculations as training data. Preferred potential (PFP) may also be used as NNP. Using NNP as the machine learning potential allows for fast and highly accurate calculation of the binding energy of the molecular structures of various compounds.

[0041] The physical indexes are an index relating to the energy required to cleave the bond between a first atom and a second atom that constitute the compound and are bonded to each other, an index relating to the charge of the atoms that constitute the compound, an index relating to the molecular weight of a first partial structure (hereinafter also referred to as a "first fragment") of the compound that includes the first atom, and an index relating to the molecular weight of a second partial structure (hereinafter also referred to as a "second fragment") of the compound that includes the second atom but does not include the first partial structure. The physical index includes at least one of these as an index.

[0042] In this embodiment, the graph database creation unit 12 calculates physical indices including an index related to energy, an index related to the charge of the atoms constituting the compound, an index related to the molecular weight of the first fragment, and an index related to the molecular weight of the second fragment.

[0043] An example of an energy index is the bond energy between a first atom and a second atom. The bond energy is a value obtained by subtracting the energies of the first and second fragments from the energy of the molecule. For example, in a compound having a molecular structure as shown in FIG. 3, when a bond is broken between one atom (black in FIG. 4) and another atom (gray in FIG. 4) bonded to the one atom in the molecule as shown in FIG. 4, the bond energy Eb is expressed as shown in the following formula (1): Bond energy = Energy of compound - (Energy of first fragment + Energy of second fragment) (1)

[0044] The first fragment is a fragment containing the first atom when the compound is cleaved between the first atom and the second atom, and in Figure 4, it consists of four atoms including one atom (black in Figure 4). The second fragment is a fragment containing the second atom when the compound is cleaved between the first atom and the second atom, and in Figure 4, it consists of two atoms including another atom (gray in Figure 4).

[0045] The charge of the atoms constituting a compound is the charge of each atom of a molecule having a molecular structure as shown in FIG. 3, as shown in FIG.

[0046] The molecular weights of the first fragment and the second fragment are the molecular weights of the respective fragments when a bond is broken between one atom (black in FIG. 6) and another atom (gray in FIG. 6) bonded to the molecule in a molecule having the molecular structure shown in FIG. 3, as shown in FIG. 6. In this case, the molecular weight of the first fragment is the molecular weight of four atoms including the one atom (black in FIG. 6), and the molecular weight of the second fragment is the molecular weight of two atoms including the other atom (gray in FIG. 6).

[0047] The graph database creation unit 12 can calculate the physical index using a graph convolutional neural network (GCNN) in which the nodes of the graph representing the structure of the compound include an index related to charge, and the edges of the graph include an index related to energy, an index related to the molecular weight of the first fragment, and an index related to the molecular weight of the second fragment.

[0048] Fig. 7 is a block diagram showing a schematic configuration of the graph database creation unit 12. As shown in Fig. 7, the graph database creation unit 12 includes a molecular structure conversion unit 121, a structure optimization calculation unit 122, a charge calculation unit 123, a bond information extraction unit 124, a partial structure creation unit (hereinafter also referred to as a "fragment creation unit") 125, a partial structure energy calculation unit (hereinafter also referred to as a "fragment energy calculation unit") 126, a bond energy calculation unit 127, a partial structure molecular weight calculation unit (hereinafter also referred to as a "fragment molecular weight calculation unit") 128, and a storage unit 129.

[0049] The molecular structure conversion unit 121 converts the molecular structure of the compound into a format that allows calculation of the machine learning potential.

[0050] The molecular structure can be converted using commonly used software, for example, the Atoms object of the ASE (Atomic Simulation Environment).

[0051] The structural optimization calculation unit 122 uses the machine learning potential to perform structural optimization calculation of the molecular structure of the compound acquired by the molecular structure conversion unit 121, and calculates the energy when the molecular structure is an optimized structure. That is, the structural optimization calculation unit 122 uses the machine learning potential to perform structural optimization of the molecular structure acquired by the molecular structure conversion unit 121, searches for a locally stable three-dimensional optimized structure (also referred to as a stable structure) of the molecular structure, and calculates the energy (Etot) of the molecular structure when it is the optimized structure (also referred to as a "structurally optimized molecular structure").

[0052] The charge calculation unit 123 uses the machine learning potential to calculate the charge of each atom constituting the molecular structure at the time of the optimized structure searched for by the structural optimization calculation unit 122, that is, the structurally optimized molecular structure.

[0053] For example, among the atoms constituting the molecular structure whose structure is optimized, the charge (Q i ) is calculated.

[0054] The charge to be calculated may be, for example, Bader charge.

[0055] Methods for calculating charge include, for example, Mulliken density analysis, Hirshfeld method, and charge equilibration method (Charge(Q) Equilibration(Eq):QEq).

[0056] Although the structural optimization calculation unit 122 and the charge calculation unit 123 use machine learning potential, either of them may not use machine learning potential.

[0057] The bond information extraction unit 124 extracts information on chemical bonds in the molecular structure at the time of the optimized structure.

[0058] Methods for extracting chemical bond information include an adjacency matrix, an order matrix, and a distance matrix.

[0059] The partial structure creation unit 125 uses the information on chemical bonds extracted by the bond information extraction unit 124 to create a first partial structure (also referred to as a first fragment) and a second partial structure (also referred to as a second fragment) when each bond between atoms constituting the structurally optimized molecular structure is cut.

[0060] The partial structure creating unit 125 may perform structural optimization on the first fragment and the second fragment.

[0061] In the molecular structure conversion unit 121, the molecular structure of the compound is converted into a format that allows for machine learning potential calculations to be performed, for example, using the Atoms object of ASE, and therefore the first fragment and second fragment whose structure is optimized in the partial structure creation unit 125 are also represented in a format that allows for machine learning potential calculations to be performed.

[0062] The partial structure energy calculation unit 126 calculates the energy of the first fragment and the second fragment created by the partial structure creation unit 125 .

[0063] For example, among the multiple bonds between atoms present in the molecular structure that has been structurally optimized, the p-side fragments (p = p1, p2) of the kth bond (k is an integer of 1 or more) are defined as the first fragment and the second fragment, and the energy (E k,p ) is calculated.

[0064] Note that p1 and p2 of p may be determined according to the value of index i of the atoms at both ends of the kth bond. That is, for the p-side fragment, one of p1 and p2 may be a first fragment whose structure is optimized, and the other may be a second fragment whose structure is optimized. For example, if the molecular weight of the first fragment is larger than that of the second fragment, p1 may be written as small and p2 as large. If the molecular weight of the first fragment is smaller than that of the second fragment, p1 may be written as large and p2 as small. If the molecular weight of the first fragment is the same as that of the second fragment, p1 may be written as equivalent.

[0065] The bond energy calculation unit 127 calculates the bond energy that bonds the atoms of the first fragment and the second fragment created by the partial structure creation unit 125 together.

[0066] For example, the bond energy (E k) is calculated by dividing the binding energy of the p-side fragment (p=p1) of the kth bond by the binding energy of the first fragment (E k,p1 ), and the binding energy of the p-side fragment (p = p2) of the kth bond is the binding energy of the second fragment (E k,p2 ) In this case, the bond energy (E k ) is calculated by subtracting the energy (E) of the first fragment from the energy (E) of the compound having the molecular structure that has been structurally optimized, which is calculated by the structural optimization calculation unit 122, as shown in the following formula (2): k,p1 ) and the energy of the second fragment (E k,p2 ) and is calculated by subtracting (E k )=Etot-(E k,p1 +E k,p2 ) ···(2)

[0067] The partial structure molecular weight calculation unit 128 calculates the molecular weights of the first fragment and the second fragment created by the partial structure creation unit 125 .

[0068] For example, the molecular weight (M k,p ) is the molecular weight of the first fragment and the second fragment. The molecular weight (M k,p ) to the molecular weight of the first fragment (M k,p1 ) and the molecular weight of the other p-side fragment (p=p2) of the kth bond (M k,p ) to the molecular weight of the second fragment (M k,p2 ) is calculated as

[0069] The storage unit 129 stores the physical indices for the atoms constituting the compound calculated based on the machine learning potential in the graph database D11. That is, the storage unit 129 stores, as physical indices, the charge of each element constituting the structure-optimized molecular structure calculated by the charge calculation unit 123, the bond energy calculated by the bond energy calculation unit 127, and the molecular weights of the first fragment and the second fragment calculated by the partial structure molecular weight calculation unit 128, in the graph database D11 as feature amounts of the graph structure representing the molecular structure of the compound created by the structure conversion unit 13.

[0070] The graph database D11, together with the structural feature database D12, is input into the training database D13. When measuring the mass spectrum of a compound, the compound is ionized. The charge of each atom constituting the molecular structure of the compound is related to the charge z of the ion on the horizontal axis (m / z) of the mass spectrum of the compound. Therefore, by inputting information on the charge of each atom constituting the molecular structure of the compound calculated by the charge calculation unit 123 from the graph database D11 into the training database D13 and using it for training by the training unit 16, the prediction accuracy of the intensity index related to the intensity of the mass spectrum of the compound can be improved.

[0071] Furthermore, the bond energy between the atoms of the first fragment and the second fragment is related to the vertical axis (intensity) of the mass spectrum of the compound. Therefore, by inputting information on the bond energy calculated by the bond energy calculation unit 127 from the graph database D11 to the learning database D13 and using it for learning by the learning unit 16, it is possible to improve the prediction accuracy of the intensity index related to the intensity of the mass spectrum of the compound.

[0072] Furthermore, the molecular weights of the first and second fragments when a compound is divided relate to the mass m of the ions on the horizontal axis (m / z) of the mass spectrum of the compound. Therefore, the molecular weights of the first and second fragments calculated by the partial structure molecular weight calculation unit 128 are input from the graph database D11 to the learning database D13, and used for learning by the learning unit 16, thereby improving the prediction accuracy of the intensity index related to the intensity of the mass spectrum of the compound.

[0073] As shown in FIG. 2, in the learning device 1, the structure conversion unit 13 converts the molecular structure of the compound acquired from the file related to the molecular structure acquired by the first acquisition unit 11 into a graph structure representing the molecular structure.

[0074] For example, in a compound having a molecular structure as shown in FIG. 3, the vertices (nodes) of the graph structure represent the atoms of the compound, and the edges of the graph structure represent the bonds between the atoms, as shown in FIG. 8.

[0075] The nodes of the graph structure are given features to represent multiple properties of atoms, such as the element type, atomic number, ionic radius, and charge calculated using machine learning potential.

[0076] The edges of the graph structure are given feature quantities including information such as the binding energy calculated by the machine learning potential, the mass of the first fragment and the mass of the second fragment, and the interatomic distance.

[0077] The feature amounts of the nodes and the feature amounts of the edges of the graph structure are expressed as vectors in matrix format.

[0078] In a graph structure, for example, when the kth bond among multiple bonds between atoms in a molecular structure is broken, the fragment containing one atom (the gray atom in Figure 8) can be called the first fragment, and the fragment containing the other atom (the white atom in Figure 8) can be called the second fragment.

[0079] The structural feature database creation unit 14 creates a structural feature database D12 having structural features of compounds acquired from a file related to molecular structures by the first acquisition unit 11. It is preferable that the structural feature database creation unit 14 stores the compounds acquired from a file related to molecular structures by the first acquisition unit 11 in the structural feature database D12 in a state where the compounds are expressed as structural feature vectors representing their structural features. The structural feature database D12 created by the structural feature database creation unit 14 is used to input data to a training database D13 used when machine learning is performed by the training unit 16.

[0080] The structural feature may be anything that indicates an index relating to the molecular structure of a compound, such as a molecular descriptor.

[0081] Examples of molecular descriptors include Morgan Fingerprint, RDkit Fingerprint, and Mordred.

[0082] The structural feature database creation unit 14 stores the created structural features in a structural feature database D12. The structural feature database D12 is input to a training database D13 together with a graph database D11. To achieve high prediction accuracy, it is important to add information that is highly correlated with the target variable as input. Structural features are also used in conventional prediction models, and are information known to be correlated with mass spectra. By inputting the created structural features from the structural feature database D12 to the training database D13 and using them for training by the training unit 16, the prediction accuracy of intensity indices related to the intensity of the mass spectrum of a compound can be improved.

[0083] The second acquiring unit 15 acquires an intensity index relating to the intensity of the mass spectrum of the compound.

[0084] The intensity index relating to the intensity of the mass spectrum of a compound may be the absolute intensity of the mass spectrum of the compound, the relative intensity of the mass spectrum of the compound, or a vector corresponding to the absolute intensity or relative intensity of the mass spectrum of the compound.

[0085] The vector corresponding to the absolute intensity or relative intensity of the mass spectrum of a compound may be, for example, a vector representing a feature graph structure that integrates information on many compounds related to mass spectra, obtained by converting the node feature vectors and edge feature vectors obtained from the graph structure and the structural feature vector through addition, pooling, combination, etc. Furthermore, the vector corresponding to the absolute intensity or relative intensity of the mass spectrum of a compound may be a vector that represents the entire compound in an appropriate matrix format by further applying nonlinear transformation using a multilayer perceptron (MLP) or the like to the integrated feature vector.

[0086] When the intensity index relating to the intensity of a mass spectrum is the intensity of the mass spectrum of a compound, it is expressed as the intensity of the mass spectrum of the compound corresponding to the mass m (m / z) of the ion relative to the charge z of the ion, for example, as shown in FIG. 9.

[0087] The second acquisition unit 15 can acquire the intensity index relating to the intensity of the mass spectrum of the compound from the molecular information database 110, similar to the first acquisition unit 11. The second acquisition unit 15 may acquire the intensity index relating to the intensity of the mass spectrum of the compound from a mass spectrum database. An example of a mass spectrum database is NIST20, a database provided by the National Institute of Standards and Technology (NIST).

[0088] Mass spectra include EI spectra and MS / MS spectra. EI spectra are mass spectra obtained by the electron ionization (EI) method, which is the most commonly used method in gas chromatography-mass spectrometry (GCMS). MS / MS spectra are mass spectra obtained using a tandem mass spectrometer, which is a system in which two mass spectrometers (MS) are connected in series.

[0089] The second acquisition unit 15 stores the acquired intensity indexes relating to the intensities of the mass spectra of the compounds in the training database D13.

[0090] The training database D13 receives as explanatory variables the physical indices of atoms constituting a compound calculated based on the machine learning potential, which are created by the graph database creation unit 12 and included in the graph database D11, and the structural features of the compound, which are created by the structural feature database creation unit 14 and stored in the structural feature database D12. The training database D13 receives as objective variables the intensity indices of the mass spectrum of the compound, which are acquired by the second acquisition unit 15. The input explanatory variables and objective variables are stored in association with each other in the training database D13.

[0091] The learning unit 16 generates a trained model M1 by learning using a learning database D13 in which explanatory variables and objective variables are associated, with the physical indicators of the atoms constituting the compound calculated based on the machine learning potential and the structural features of the compound as explanatory variables and the intensity indicators of the intensity of the mass spectrum of the compound as objective variables.

[0092] In addition, the learning unit 16 may not use the structural features of the compounds contained in the structural feature database D12 as explanatory variables, but may use only the physical indicators related to the atoms constituting the compounds calculated based on the machine learning potential as explanatory variables.

[0093] The output unit 17 outputs, by display or the like, information on the training database D13 used in training the trained model M1, and information on the trained model M1.

[0094] As described above, the learning device 1 includes a learning unit 16, which generates a trained model M1 using a training database D13 in which physical indices for atoms constituting a compound calculated based on a machine learning potential are used as explanatory variables and intensity indices for the intensities of the mass spectra of the compound are used as objective variables. The physical indices for atoms used in the training database D13 are calculated using a machine learning potential and therefore can be obtained with high accuracy. In particular, when the physical indices are calculated using a highly accurate machine learning potential, the physical indices can be obtained with even higher accuracy. Therefore, the learning device 1 can generate a trained model M1 that can improve the prediction accuracy when predicting intensity indices for the intensities of the mass spectra of various compounds by training using the training database D13 containing highly accurate physical indices in the learning unit 16.

[0095] Furthermore, the trained model M1 can predict the intensity indexes related to the mass spectrum intensities of compounds other than those having the same structure as the compounds stored in the training database D13. Therefore, the learning device 1 can generate a trained model M1 with high versatility.

[0096] Furthermore, the trained model M1 can be used to predict intensity indices related to the intensity of the mass spectrum of a compound having an unknown structure. Thus, the learning device 1 can generate a trained model M1 that enables even a non-expert to predict the mass spectrum of a compound having an unknown structure and predict and analyze the molecular structure of the compound.

[0097] In the learning device 1, the physical index preferably includes at least one index selected from the group consisting of an index related to energy, an index related to the charge of atoms constituting the compound, an index related to the molecular weight of the first fragment, and an index related to the molecular weight of the second fragment. By including any of the above indexes as a physical index in the learning database D13, the trained model M1 can further improve the prediction accuracy of the intensity index related to the intensity of the mass spectrum of the compound. Therefore, the learning device 1 can generate a trained model M1 that can further improve the prediction accuracy when predicting an index related to the intensity of the mass spectrum.

[0098] In the learning device 1, the physical indices are preferably calculated using GCNN, in which nodes of a graph representing the structure of a compound include an index related to charge, and edges of the graph include an index related to energy, an index related to the molecular weight of a first fragment, and an index related to the molecular weight of a second fragment. By using the physical indices calculated using GCNN in the learning database D13, the trained model M1 can further improve the prediction accuracy of intensity indices related to the intensity of a compound's mass spectrum. Therefore, the learning device 1 can generate a trained model M1 that can further improve the prediction accuracy when predicting indices related to the intensity of a mass spectrum.

[0099] In the learning device 1, the training database D13 preferably includes structural features of compounds as explanatory variables, and the physical indices and structural features of the compounds are associated with intensity indices related to the intensities of the mass spectra of the compounds. That is, the learning unit 16 preferably generates the trained model M1 using a training database D13 in which the physical indices related to the atoms constituting the compounds calculated based on the machine learning potential and the structural features of the compounds are associated with the intensity indices related to the intensities of the mass spectra of the compounds as objective variables. In this way, the learning device 1 can generate a trained model M1 that can further improve the prediction accuracy when predicting intensity indices related to the intensities of the mass spectra of various compounds by training the learning unit 16 using the training database D13 including highly accurate physical indices and structural features of the compounds.

[0100] The learning device 1 does not need to have all of the above components, and may be configured to have only the necessary components depending on the application.

[0101] <Prediction device> The prediction device according to the present embodiment predicts an intensity index related to the intensity of the mass spectrum of a compound to be predicted as a response variable by inputting the physical indexes of the compound to be predicted and the structural feature quantities of the compound to be predicted as explanatory variables into a trained model trained using a training database in which physical indexes related to the atoms constituting the compound, calculated based on a machine learning potential, are associated with intensity indexes related to the intensity of the mass spectrum of the compound.

[0102] In the prediction device according to this embodiment, the node feature used as the physical index for the atoms constituting the compound to be predicted is the atomic charge calculated using the machine learning potential, and the edge feature is the bond energy calculated using the machine learning potential (see FIG. 1(a)). Therefore, the node and edge feature values ​​have high accuracy. The prediction device according to this embodiment inputs the physical index of the compound to be predicted as an explanatory variable into a trained model generated using a training database in which physical indexes for the atoms constituting the compound are associated with intensity indexes related to the intensity of the mass spectrum of the compound. This allows the prediction device according to this embodiment to improve the prediction accuracy when predicting intensity indexes related to the intensity of the mass spectrum of the compound to be predicted.

[0103] Fig. 10 is a block diagram showing a schematic configuration of a prediction device according to this embodiment. As shown in Fig. 10, the prediction device 2A includes an acquisition unit 21, a graph data creation unit 22, a structural conversion unit 23, a structural feature data creation unit 24, a prediction unit 25, an output unit 26, and a trained model M2. The prediction device 2A uses the trained model M2 to predict an intensity index related to the intensity of a molecular mass spectrum.

[0104] The prediction device 2A does not need to have all of the above components, and may be configured to have only the necessary components depending on the application.

[0105] The acquisition unit 21 acquires, as a prediction target, a file relating to the molecular structure of the compound to be predicted from a molecular information database 210. The molecular information database 210 may be one created by the user or an existing database such as PubChem.

[0106] The compound to be predicted is a compound whose mass spectrum is to be predicted.

[0107] The molecular information database 210 contains files related to molecular structures, similar to the molecular information database 110 of the learning device 1 according to the present embodiment described above. The acquisition unit 21 is similar to the first acquisition unit 11 of the learning device 1 according to the present embodiment described above, except that the target to be acquired is a compound to be predicted, and therefore details thereof will be omitted.

[0108] The graph data creation unit 22 creates graph data based on physical indicators of the atoms constituting the compound to be predicted, calculated using machine learning potentials based on a file relating to the molecular structure of the compound to be predicted acquired by the acquisition unit 21.

[0109] The physical indexes are similar to the physical indexes calculated by the graph database creation unit 12 of the learning device 1 according to the present embodiment described above, and therefore details thereof will be omitted.

[0110] The structure conversion unit 23 converts the file relating to the molecular structure of the compound to be predicted, acquired by the acquisition unit 21, into a graph structure representing the acquired molecular structure of the compound to be predicted. The structure conversion unit 23 is similar to the structure conversion unit 13 of the learning device 1 according to the present embodiment described above, except that the target to be converted is the compound to be predicted, and therefore details thereof will be omitted.

[0111] The structural feature data creation unit 24 creates structural feature data having the structural feature of the compound to be predicted, from the file related to the molecular structure acquired by the acquisition unit 21. The structural feature is similar to the structural feature created by the structural feature database creation unit 14 of the learning device 1 according to the present embodiment described above, and therefore details thereof will be omitted.

[0112] The prediction unit 25 predicts an intensity index related to the intensity of the mass spectrum of the compound to be predicted by inputting the physical indexes related to the atoms constituting the compound to be predicted, created by the graph data creation unit 22, and the structural features of the compound to be predicted, created by the structural feature data creation unit 24, into the trained model M2.

[0113] The trained model M2 is a trained model trained using a training database in which physical indicators related to the atoms constituting the molecule and structural features of the molecule calculated based on the machine learning potential are associated with intensity indicators related to the intensity of the mass spectrum of the molecule, in order to use the trained model M1 of the learning device 1 according to the present embodiment described above.

[0114] The output unit 26 outputs, by display or the like, the intensity index relating to the intensity of the mass spectrum of the compound to be predicted, which is predicted by the trained model M2 in the prediction unit 25, as a response variable.

[0115] The prediction device 2A includes a prediction unit 25. The prediction unit 25 uses a trained model M2 to predict an intensity index related to the intensity of the mass spectrum of a compound to be predicted, based on physical indices related to the atoms constituting the compound to be predicted and structural features of the compound to be predicted, which are calculated based on the machine learning potential. The trained model M2 uses the trained model M1 generated by the learning device 1 according to the present embodiment. The physical indices related to the atoms of the compound to be predicted used in the prediction unit 25 are calculated using the machine learning potential, and therefore can be obtained with high accuracy. Therefore, by using the trained model M2 and highly accurate physical indices and structural features of the compound to be predicted in the prediction unit 25, the prediction device 2A can improve the prediction accuracy of the intensity index related to the intensity of the mass spectrum of the compound to be predicted, based on the physical indices related to the atoms constituting the compound to be predicted and the structural features of the compound to be predicted. Therefore, the prediction device 2A can improve the prediction accuracy when predicting intensity indices related to the intensity of the mass spectrum of various compounds.

[0116] Therefore, the prediction device 2A can quickly and accurately predict the structure of an unknown compound to be predicted.

[0117] When predicting intensity indices related to the mass spectrum intensity of a compound, simply inputting information about the molecular structure as explanatory variables and using GCNN, multilayer perceptrons (MLPs), or the like to predict the mass spectrum intensity of the compound generally results in a large deviation from the experimental value, resulting in low prediction accuracy. To improve the prediction accuracy of the mass spectrum intensity of a compound, a method is used in which, in addition to structural information, quantities correlated with the mass spectrum, such as binding energy, are input. However, the binding energy is calculated using classical interatomic potentials, resulting in low calculation accuracy and making it difficult to predict the mass spectrum. Furthermore, quantum chemical calculations require high computational costs, resulting in significant financial and time-consuming efforts for prediction. The prediction device 2A can use the trained model M2 to improve the prediction accuracy when predicting intensity indices related to the mass spectrum intensity of the target compound, thereby reducing the financial and time-consuming efforts required for predicting intensity indices related to the mass spectrum intensity.

[0118] Furthermore, by using the trained model M2, the prediction device 2A can predict the intensity index related to the intensity of the mass spectrum of the compound to be predicted, even for compounds other than those having the same structure as the compound to be predicted stored in the training database D13 (see FIG. 2) used when generating the trained model M1 (see FIG. 2) used in the trained model M2. Therefore, the prediction device 2A can have high versatility.

[0119] Furthermore, the prediction device 2A can predict intensity indices related to the intensity of the mass spectrum of a compound having an unknown structure. Therefore, the prediction device 2A allows even a non-expert to predict the mass spectrum of a compound having an unknown structure and predict and analyze the molecular structure of the compound.

[0120] The learning device 1 and prediction device 2A may be configured as a learning system and a prediction system. That is, the learning device 1 is configured as a standalone device such as a PC (Personal Computer) that includes each component within the device, but one or more of the components may be located outside the device and connected via a network.

[0121] For example, the learning database may be provided on a cloud. In this case, the learning device 1 is configured as a learning system by the learning database connected via a network.

[0122] Similarly, one or more of the components of the prediction device 2A may be arranged outside the device and connected via a network.

[0123] <Hardware configuration of the learning device and prediction device> Next, an example of the hardware configuration of the learning device 1 and the prediction device 2A will be described. FIG. 11 is a block diagram showing the hardware configuration of the learning device 1 and the prediction device 2A. As shown in FIG. 11, the learning device 1 and the prediction device 2A are configured as information processing devices (computers), and can be physically configured as a computer system including a processor 101 as an arithmetic processing unit, a memory 102, an input device 103 as an input device, an output device 104, a communication module 105, and an auxiliary storage device 106 such as a hard disk. These are connected to each other by a bus 107. Note that the output device 104 and the auxiliary storage device 106 may be provided externally.

[0124] Processor 101 controls the overall operation of learning device 1 and prediction device 2A and performs various information processing. Processor 101 can predict intensity indices related to the intensity of the mass spectrum of a compound by executing, for example, a learning method and a prediction method or a learning program and a prediction program (described below) stored in memory 102 or auxiliary storage device 106. As processor 101, for example, a CPU (Central Processing Unit) or the like is used.

[0125] The memory 102 includes a RAM (Random Access Memory) and a ROM (Read Only Memory) that are main storage devices, and may store a learning program and a prediction program.

[0126] The RAM may be used as a work area for the processor 101 and may include non-volatile RAM for storing key control parameters and information.

[0127] The ROM stores a basic input / output program, etc. The learning program and the prediction program may be stored in the ROM.

[0128] The input device 103 is an input device such as a keyboard, a mouse, operation buttons, a touch panel, or a display screen, and receives information input by a user as an instruction signal, and outputs the instruction signal to the processor 101 .

[0129] The output device 104 is a display device such as a monitor display, a speaker, a printing device such as a printer, etc. In the output device 104, for example, information such as the learning results and the predicted results of the oil film density are displayed on a display device such as a monitor display, and the displayed screen is updated in response to input operations via the input device 103 or the communication module 105.

[0130] The communication module 105 is a data transmission / reception device such as a network card, and functions as a communication interface that receives information from an external data recording server or the like and outputs analysis information to other electronic devices.

[0131] The auxiliary storage device 106 is a storage device such as an SSD (Solid State Drive) or an HDD (Hard Disk Drive), and stores, for example, various data, files, etc. required for the operation of the learning device 1 and the prediction device 2A.

[0132] Each function of the learning device 1 and the prediction device 2A is realized by reading predetermined computer software (including a learning program and a prediction program) from the memory 102 or the auxiliary storage device 106 and executing it with the processor 101, thereby reading and writing data in the memory 102, the auxiliary storage device 106, etc., and operating the input device 103, the output device 104, and the communication module 105.

[0133] Therefore, each part of the learning device 1 and the prediction device 2A shown in Figures 2 and 10 is realized by software and hardware working together in a computer equipped with the learning device 1 and the prediction device 2A, where the processor 101 executes predetermined computer software (including a learning program and a prediction program) that is pre-stored.

[0134] A computer program that implements at least some of the functions of the parts of the learning device 1 and the prediction device 2A shown in Figures 2 and 10 may be installed in the storage of one or more computers. The processor 101 of one or more computers may load the installed computer program into a main memory and execute it to fulfill the functions of the parts of the learning device 1 and the prediction device 2A shown in Figures 2 and 10.

[0135] The learning device 1 and the prediction device 2A shown in Figures 2 and 10 may be realized by one or more processors 101. Here, the processor 101 may refer to one or more electronic circuits arranged on one chip, or to one or more electronic circuits arranged on two or more chips or two or more devices. When multiple electronic circuits are used, the respective electronic circuits may communicate with each other by wire or wirelessly.

[0136] 2 and 10 may be executed by a single computer or may be executed in a distributed manner by multiple computers. When the functions of the learning device 1 and prediction device 2A shown in FIGS. 2 and 10 are executed in a distributed manner by multiple computers, the multiple computers may transmit and receive data via a communication network including a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or the Internet.

[0137] The learning program and prediction program can be stored, for example, in the memory 102 or auxiliary storage device 106 of the computer. Alternatively, the learning program and prediction program can be stored on a computer connected to a communication line such as the Internet, and some or all of the learning program and prediction program can be provided by being downloaded via the communication line. Furthermore, the learning program and prediction program can be configured to be provided or distributed via the communication line.

[0138] The learning program and prediction program may be recorded (including installed) into a computer from a state in which part or all of them are stored on a portable storage medium such as an optical disk such as a CD-ROM or DVD-ROM, or a semiconductor memory such as a flash memory.

[0139] <Learning Method> Next, a learning method according to this embodiment will be described. The learning method according to this embodiment can be performed using the above-described learning device 1. Therefore, in each step, some of the content already explained for the above-described learning device 1 will be omitted.

[0140] Fig. 12 is a flowchart illustrating the learning method according to this embodiment. As shown in Fig. 12, in the learning method according to this embodiment, the first acquisition unit 11 acquires a file related to the molecular structure of a compound from the molecular information database 110 (first acquisition step: step S11).

[0141] Next, the structure conversion unit 13 converts the molecular structure of the compound acquired from the file relating to the molecular structure acquired in the first acquisition step S11 into a graph structure representing the molecular structure (structure conversion step: step S12).

[0142] Next, the graph database creation unit 12 creates a graph database D11 having physical indices for the atoms that constitute the compound calculated using the machine learning potential based on the file related to the molecular structure acquired in the first acquisition step S11 (graph database creation step: step S13).

[0143] Fig. 13 is a flowchart illustrating the graph database creation step S13. As shown in Fig. 13, in the graph database creation step S13, the molecular structure conversion unit 121 converts the molecular structure of the compound into a format that can execute the calculation of the machine learning potential (molecular structure conversion step: step S131).

[0144] Next, the structural optimization calculation unit 122 uses the machine learning potential to perform a structural optimization calculation of the molecular structure of the compound acquired by the molecular structure conversion unit 121, and calculates the energy when the molecular structure is the optimized structure (structural optimization calculation process: step S132).

[0145] Next, the charge calculation unit 123 uses the machine learning potential to calculate the charge of each atom constituting the molecular structure at the time of the optimized structure searched for in the structural optimization calculation step S132, i.e., the structurally optimized molecular structure (charge calculation step: step S133).

[0146] In the graph database creation step S13, the bond information extraction unit 124 extracts information on chemical bonds of the molecular structure in the optimized structure searched for in the structure optimization calculation step S132 (bond information extraction step: step S134).

[0147] Next, the partial structure creation unit 125 uses the information on chemical bonds extracted in the bond information extraction step S134 to create first and second fragments when each bond between atoms constituting the structurally optimized molecular structure is cut (partial structure creation step: step S135).

[0148] Next, the partial structure energy calculation unit 126 calculates the energy of the first fragment and the second fragment created in the partial structure creation step S135 (partial structure energy calculation step: step S136).

[0149] Next, the bond energy calculation unit 127 calculates the bond energy that bonds the atoms of the first fragment and the second fragment created in the partial structure creation step S135 (bond energy calculation step: step S137).

[0150] In the graph database creation step S13, the partial structure molecular weight calculation unit 128 calculates the molecular weights of the first fragment and the second fragment created in the partial structure creation step S135 (partial structure molecular weight calculation step: step S138).

[0151] Next, the storage unit 129 stores the physical indices for the atoms that constitute the compound calculated based on the machine learning potential in the graph database D11 (storage step: step S139).

[0152] That is, the storage unit 129 stores the charges of each element constituting the structurally optimized molecular structure calculated in the charge calculation step S133, the bond energy bonding the atoms of the first fragment and the second fragment calculated in the bond energy calculation step S137, and the molecular weights of the first fragment and the second fragment calculated in the molecular weight calculation step S138 of the partial structure as physical indicators for the atoms constituting the compound calculated based on the machine learning potential in the graph database D11 as feature quantities of the graph structure representing the molecular structure of the compound created in the structure conversion step S12.

[0153] In the flow of the graph database creation step S13 shown in FIG. 13, the charge calculation step S133 is performed before the bond information extraction step S134, and the bond energy calculation step S137 is performed before the partial structure molecular weight calculation step S138, but the order of each step is not particularly limited.

[0154] The charge calculation step S133 and the bond information extraction step S134 may be performed simultaneously, or the bond information extraction step S134 may be performed before the charge calculation step S133.

[0155] Furthermore, the bond energy calculation step S137 and the partial structure molecular weight calculation step S138 may be performed simultaneously, or the partial structure molecular weight calculation step S138 may be performed before the bond energy calculation step S137.

[0156] Next, as shown in FIG. 12, in the learning method according to this embodiment, the structural feature database creation unit 14 creates a structural feature database D12 having the structural features of the compound acquired from the file related to the molecular structure in the first acquisition step S11 (structural feature database creation step: step S14).

[0157] Next, the second acquisition unit 15 acquires an intensity index relating to the intensity of the mass spectrum of the compound (second acquisition step: step S15).

[0158] Next, the graph database D11 created in the graph database creation process S13, the structural feature database D12 created in the structural feature database creation process S14, and the intensity index related to the mass spectrum intensity of the compound acquired in the second acquisition process S15 are input into the training database D13 to create the training database D13 (training database creation process: step S16).

[0159] The training database D13 receives as explanatory variables the physical indices of atoms constituting the compound calculated based on the machine learning potential, which are included in the graph database D11, and the structural features of the compound stored in the structural feature database D12. The training database D13 receives as objective variables the intensity indices of the mass spectrum of the compound. The input explanatory variables and objective variables are stored in association with each other in the training database D13.

[0160] Next, the learning unit 16 generates a trained model M1 by learning using a training database D13 in which explanatory variables and objective variables are associated, with the physical indicators for the atoms constituting the compound calculated based on the machine learning potential and the structural features of the compound as explanatory variables and the intensity indicator for the intensity of the mass spectrum of the compound as objective variables (learning process: step S17).

[0161] Next, the output unit 17 outputs, by display or the like, information on the training database D13 used in training the trained model M1, information on the trained model M1, and the like (output step: step S18).

[0162] The training method according to this embodiment includes a training step S16, in which a trained model M1 is generated using a training database D13 in which physical indices for atoms constituting a compound calculated based on the machine learning potential are used as explanatory variables and intensity indices for the intensities of the mass spectrum of the compound are used as objective variables. The physical indices for atoms used in the training database D13 are calculated using the machine learning potential and therefore can be obtained with high accuracy. In particular, when the physical indices are calculated using a highly accurate machine learning potential, the physical indices can be obtained with even higher accuracy. Therefore, the training method according to this embodiment trains using the training database D13 containing highly accurate physical indices in the training step S16, thereby generating a trained model M1 that can improve the prediction accuracy when predicting intensity indices for the intensities of the mass spectra of various compounds.

[0163] <Prediction method> Next, a prediction method according to this embodiment will be described. The prediction method according to this embodiment can be performed using the prediction device 2A described above. Therefore, in each step, some of the content already described for the prediction device 2A will be omitted.

[0164] Fig. 14 is a flowchart showing the prediction method according to this embodiment. As shown in Fig. 14, in the prediction method according to this embodiment, the acquisition unit 21 acquires, as a prediction target, a file related to the molecular structure of a compound to be predicted from the molecular information database 210. (Acquisition process: step S21).

[0165] Next, the structure conversion unit 23 converts the file relating to the molecular structure of the compound to be predicted, acquired in the acquisition step S21, into a graph structure representing the acquired molecular structure of the compound to be predicted (structure conversion step: step S22).

[0166] Next, the graph data creation unit 22 creates graph data based on the physical indicators of the atoms constituting the compound to be predicted, calculated using the machine learning potential based on the file related to the molecular structure of the compound to be predicted acquired in the acquisition process S21 (graph data creation process: step S23).

[0167] The method for creating the physical indexes is the same as the method for creating the physical indexes calculated in the graph database creation step S13 in the learning method according to this embodiment described above, and therefore details thereof will be omitted.

[0168] Next, the structural feature data creation unit 24 creates structural feature data having the structural feature of the compound to be predicted from the file related to the molecular structure acquired in the acquisition step S21 (structural feature data creation step: step S24).

[0169] Next, the prediction unit 25 inputs the physical indices related to the atoms constituting the compound to be predicted, created in the graph data creation process S23, and the structural features of the compound to be predicted, created in the structural feature data creation process S24, into the trained model M2, thereby predicting an intensity index related to the intensity of the mass spectrum of the compound to be predicted (prediction process: step S25).

[0170] Next, the output unit 26 outputs, by display or the like, the intensity index relating to the intensity of the mass spectrum of the compound to be predicted, predicted by the trained model M2 in the prediction step S25, as the objective variable (output step: step S26).

[0171] The prediction method according to this embodiment includes a prediction step S25. In the prediction step S25, a trained model M2 is used to predict an intensity index related to the intensity of the mass spectrum of the compound to be predicted, based on the physical indices related to the atoms constituting the compound to be predicted and the structural feature values ​​of the compound to be predicted, which are calculated based on the machine learning potential. The trained model M2 uses the trained model M1 generated by the learning method according to this embodiment. The physical indices related to the atoms of the compound to be predicted used in the prediction step S25 are calculated using the machine learning potential, and therefore can be obtained with high accuracy. Therefore, the prediction method according to this embodiment can improve the prediction accuracy of the intensity index related to the intensity of the mass spectrum of the compound to be predicted, based on the physical indices related to the atoms constituting the compound to be predicted and the structural feature values ​​of the compound to be predicted, by using the trained model M2 and highly accurate physical indices and structural feature values ​​of the compound to be predicted in the prediction step S25. Therefore, the prediction method according to this embodiment can improve the prediction accuracy when predicting intensity indices related to the intensity of the mass spectrum of various compounds.

[0172] Therefore, by using the prediction method according to this embodiment, the structure of an unknown compound to be predicted can be predicted quickly and with high accuracy.

[0173] The prediction method of this embodiment can improve the prediction accuracy when predicting intensity indices related to the intensity of the mass spectrum of the compound to be predicted using the trained model M2, thereby reducing the cost and time burden required to predict intensity indices related to the intensity of the mass spectrum.

[0174] Furthermore, by using the trained model M2, the prediction method according to this embodiment can predict the intensity index related to the intensity of the mass spectrum of a compound to be predicted, even for compounds other than those having the same structure as the compound to be predicted stored in the training database D13 (see FIG. 2) used when generating the trained model M1 (see FIG. 2) used in the trained model M2. Therefore, the prediction method according to this embodiment can be highly versatile.

[0175] Furthermore, the prediction method according to this embodiment can predict intensity indices related to the intensity of the mass spectrum of a compound having an unknown structure. Therefore, the prediction method according to this embodiment allows even a non-expert to predict the mass spectrum of a compound having an unknown structure and predict and analyze the molecular structure of the compound.

[0176] [Second embodiment] A learning device and a prediction device according to an embodiment of the present invention will be described. Note that in this embodiment, the prediction device according to this embodiment is similar to the prediction device 2A according to the second embodiment, except that the prediction device creates a database for prediction, similar to the learning database D13 of the learning device 1 according to the first embodiment, and therefore only the prediction device will be described.

[0177] <Prediction device> The prediction device according to this embodiment uses files related to the molecular structures of multiple compounds to be predicted, in the same manner as in the creation of the training database D13 of the learning device 1 according to the first embodiment described above, to create a prediction database in which the physical indices related to the atoms constituting the compound to be predicted, calculated based on the machine learning potential, and the structural feature quantities of the compound to be predicted, are used as explanatory variables, and an intensity index related to the intensity of the mass spectrum of the molecule to be predicted is used as the objective variable.The prediction device according to this embodiment then uses the created prediction database to input the physical indices of each compound to be predicted as explanatory variables into the trained model M2, thereby making predictions using the intensity index related to the intensity of the mass spectrum of each compound to be predicted as the objective variable.

[0178] Fig. 15 is a block diagram showing another example of the schematic configuration of a prediction device according to this embodiment. As shown in Fig. 15, the prediction device 2B includes an acquisition unit 21, a graph database creation unit 27, a structure conversion unit 23, a structural feature database creation unit 28, a prediction unit 25, an output unit 26, a graph database D21, a structural feature database D22, a prediction database D23, and a trained model M2.

[0179] The graph database D21, the structural feature database D22, the prediction database D23, and the trained model M2 are similar to the graph database D11, the structural feature database D12, the training database D13, and the trained model M1 of the training device 1 according to the first embodiment described above, and therefore details of these will be omitted.

[0180] The graph database creation unit 27 creates a graph database D21 having physical indices related to the atoms that make up the compound calculated using the machine learning potential from the file related to the molecular structure of the compound to be predicted acquired by the acquisition unit 21. The graph database creation unit 27 is similar to the graph database creation unit 12 of the learning device 1 according to the first embodiment described above, except that the object to be created is the compound to be predicted, and therefore details thereof will be omitted.

[0181] The structural feature database creation unit 28 creates a structural feature database D22 having structural features of the acquired compound from a file related to the molecular structure of the compound to be predicted acquired by the acquisition unit 21. The structural feature database creation unit 28 is similar to the structural feature database creation unit 14 of the learning device 1 according to the first embodiment described above, except that the target is the compound to be predicted, and therefore details thereof will be omitted.

[0182] The prediction database D23 receives as explanatory variables the physical indicators of the atoms constituting the compound to be predicted, which are calculated based on the machine learning potential and are created by the graph database creation unit 27 and included in the graph database D21, and the structural features of the compound to be predicted, which are created by the structural feature database creation unit 28 and stored in the structural feature database D22, and stores these inputs as explanatory variables.

[0183] The prediction unit 25 predicts an intensity index related to the intensity of the mass spectrum of the molecule to be predicted by inputting the physical index and structural features of the molecule to be predicted, which are stored in the prediction database D23, into the trained model M2.

[0184] The prediction device 2B includes a structure conversion unit 23, a graph database creation unit 27, a structural feature database creation unit 28, and a prediction unit 25. The physical indices of atoms in the prediction target compound stored in the prediction database D4 are calculated using machine learning potentials in the graph database creation unit 27, and thus can be obtained with high accuracy. Therefore, when using multiple prediction target compounds, the prediction device 2B creates a prediction database D4 in advance using the multiple prediction target compounds, and then uses the created prediction database D4 to, for each prediction target compound, use a prediction database D23 and a trained model M2 containing highly accurate physical indices in the prediction unit 25. This allows the prediction unit 25 to improve the prediction accuracy of intensity indices related to the mass spectral intensities of each prediction target compound from the physical indices related to the atoms constituting each prediction target compound. Therefore, like the prediction device 2A according to the first embodiment described above, the prediction device 2B can improve the prediction accuracy when predicting intensity indices related to the mass spectral intensities of various compounds.

[0185] <Prediction method> Next, a prediction method according to this embodiment will be described. The prediction method according to this embodiment can be performed using the prediction device 2B described above. Therefore, in each step, some of the content already described for the prediction device 2B will be omitted.

[0186] Fig. 16 is a flowchart showing the prediction method according to this embodiment. As shown in Fig. 16, in the prediction method according to this embodiment, the acquisition unit 21 acquires, as a prediction target, a file related to the molecular structure of a compound to be predicted from the molecular information database 210 (acquisition step: step S31).

[0187] Next, the structure conversion unit 23 converts the file relating to the molecular structure of the compound to be predicted, acquired in the acquisition step S31, into a graph structure representing the acquired molecular structure of the compound to be predicted (structure conversion step: step S32).

[0188] Next, the graph database creation unit 27 creates a graph database D21 having physical indicators for the atoms constituting the compound to be predicted calculated using the machine learning potential based on the file related to the molecular structure of the compound to be predicted acquired in the acquisition process S31 (graph database creation process: step S33).

[0189] The graph database creation step S33 is similar to the graph database creation step S13 shown in FIG. 13 of the learning method according to the first embodiment described above, except that the target is a compound to be predicted, and therefore details will be omitted.

[0190] The graph database creation unit 27 stores the physical indices of the atoms constituting the compound to be predicted as feature quantities of the graph structure representing the molecular structure of the compound created in the structure conversion step S32.

[0191] Next, the structural feature database creation unit 28 creates a structural feature database D22 having the structural features of the compound to be predicted, from the file related to the molecular structure of the compound to be predicted acquired in the acquisition step S21 (structural feature database creation step: step S34).

[0192] Next, the graph database D21 created in the graph database creation process S33 and the structural feature database D22 created in the structural feature database creation process S34 are combined to create a prediction database D23 (prediction database creation process: step S35).

[0193] The prediction database D23 receives as explanatory variables the physical indices of the atoms constituting the compound to be predicted, calculated based on the machine learning potential and stored in the graph database D21 created in the graph database creation step S33, and the structural features of the compound to be predicted, stored in the structural feature database D22 created in the structural feature database creation step S34, and stores these as explanatory variables.

[0194] Next, the prediction unit 25 inputs the physical indices related to the atoms constituting the compound to be predicted and the structural features of the compound to be predicted, which are stored in the prediction database D23, into the trained model M2, thereby predicting an intensity index related to the intensity of the mass spectrum of the compound to be predicted (prediction process: step S36).

[0195] Next, the output unit 26 outputs, by display or the like, the intensity index relating to the intensity of the mass spectrum of the compound to be predicted, predicted by the trained model M2 in the prediction step S36, as the objective variable (output step: step S37).

[0196] The prediction method according to this embodiment includes a structure conversion step S32, a graph database creation step S33, a structural feature database creation step S34, and a prediction step S36. The physical indices of the atoms of the target compound stored in the prediction database D23 are calculated using machine learning potentials in the graph database creation step S33, and thus can be obtained with high accuracy. Therefore, the prediction method according to this embodiment creates a prediction database D4 in advance using multiple target compounds, and then, for each target compound, uses the created prediction database D4 to predict the prediction database D23 and the trained model M2 containing highly accurate physical indices in the prediction step S36. This allows the prediction step S36 to improve the prediction accuracy of intensity indices related to the mass spectral intensities of each target compound based on the physical indices related to the atoms constituting each target compound. Therefore, the prediction method according to this embodiment also improves the prediction accuracy when predicting intensity indices related to the mass spectral intensities of various compounds.

[0197] Although the embodiments have been described above, they are presented as examples and the present invention is not limited to the above embodiments. The above embodiments can be implemented in various other forms, and various combinations, omissions, substitutions, modifications, etc. can be made without departing from the spirit of the invention. These embodiments and their modifications are included within the scope and spirit of the invention, and are also included in the scope of the invention and its equivalents as set forth in the claims.

[0198] The embodiments of the present invention are as follows, for example. <1> A learning device comprising a learning unit that generates a trained model by learning using a learning database in which physical indices related to atoms constituting a compound calculated based on machine learning potential are associated with intensity indices related to the intensity of the mass spectrum of the compound. <2> the physical index includes at least one index selected from the group consisting of an index relating to energy for cleaving a bond between a first atom and a second atom that constitute the compound and are bonded to each other, an index relating to the charge of the atom that constitutes the compound, an index relating to the molecular weight of a first partial structure of the compound that includes the first atom, and an index relating to the molecular weight of a second partial structure of the compound that includes the second atom but does not include the first partial structure; <1> The learning device according to claim 1. <3> the physical index is calculated using a graph convolutional neural network in which nodes of a graph representing a structure of the compound include the index related to the charge, and edges of the graph include the index related to the energy, an index related to the molecular weight of the first partial structure, and an index related to the molecular weight of the second partial structure. <2> The learning device according to claim 1. <4> the training database includes structural features of the compounds, the physical index and the structural feature of the compound are associated with an intensity index relating to the intensity of the mass spectrum of the compound; <1> ~ <3> 10. The learning device according to claim 9, wherein <5> A prediction device comprising a prediction unit that predicts an intensity index related to the intensity of a mass spectrum of a compound to be predicted by inputting the physical index of the compound to be predicted and structural features of the compound to be predicted into a trained model that has been trained using a training database in which physical indexes related to atoms constituting the compound, calculated based on machine learning potential, are associated with intensity indexes related to the intensity of the mass spectrum of the compound. <6> The computer A learning method that performs a learning process to generate a trained model by learning using a learning database in which physical indicators related to the atoms that constitute a compound, calculated based on machine learning potential, are associated with intensity indicators related to the intensity of the mass spectrum of the compound. <7> On the computer, A learning program that executes a learning process to generate a trained model by learning using a learning database in which physical indicators related to the atoms that constitute a compound calculated based on machine learning potential are associated with intensity indicators related to the intensity of the mass spectrum of the compound. <8> The computer A prediction method comprising: executing a prediction step of predicting the intensity index relating to the intensity of the mass spectrum of a compound to be predicted by inputting the physical index of the compound to be predicted and structural features of the compound to be predicted into a trained model trained using a training database in which physical indexes relating to the atoms constituting the compound, calculated based on machine learning potential, and intensity indexes relating to the intensity of the mass spectrum of the compound are associated with each other. <9> On the computer, A prediction program that executes a prediction step of predicting the intensity index related to the intensity of the mass spectrum of a compound to be predicted by inputting the physical index of the compound to be predicted and structural features of the compound to be predicted into a trained model that has been trained using a training database in which physical indexes related to the atoms that constitute the compound, calculated based on machine learning potential, and intensity indexes related to the intensity of the mass spectrum of the compound are associated with each other. [Explanation of symbols]

[0199] 1 Learning device 2A, 2B prediction device 11 First acquisition part 12, 27 Graph database creation section 13, 23 Structural transformation section 14, 28 Structural feature database creation section 15 Second acquisition part 16 Learning Department 17, 26 Output section 21 Acquisition Department 22 Creating data for graphs 24. Structural feature data creation section 25 Prediction Department 121 Molecular Structure Conversion Unit 122 Structural Optimization Calculation Unit 123 Charge calculation section 124 Combined information extraction part 125 Substructure Creation Unit (Fragment Creation Unit) 126 Substructure energy calculation part (fragment energy calculation part) 127 Binding Energy Calculation Unit 128 Molecular weight calculation of partial structure (fragment molecular weight calculation) 129 Preservation Department D11, D21 Graph database D12, D22 structural feature database D13 Learning Database D23 Prediction Database M1, M2 trained models

Claims

1. A learning device comprising a learning unit that generates a trained model by learning using a learning database in which physical indices related to atoms constituting a compound calculated based on machine learning potential are associated with intensity indices related to the intensity of the mass spectrum of the compound.

2. The learning device of claim 1, wherein the physical indicators include at least one indicator selected from the group consisting of an indicator related to the energy required to cleave a bond between a first atom and a second atom that constitute the compound and are bonded to each other, an indicator related to the charge of the atoms that constitute the compound, an indicator related to the molecular weight of a first partial structure of the compound that includes the first atom, and an indicator related to the molecular weight of a second partial structure of the compound that includes the second atom but does not include the first partial structure.

3. 3. The learning device according to claim 2, wherein the physical index is calculated using a graph convolutional neural network in which nodes of a graph representing a structure of the compound include the index related to the charge, and edges of the graph include the index related to the energy, an index related to the molecular weight of the first partial structure, and an index related to the molecular weight of the second partial structure.

4. the training database includes structural features of the compounds, The learning device according to claim 1 , wherein the physical index and the structural feature amount of the compound are associated with an intensity index relating to the intensity of the mass spectrum of the compound.

5. A prediction device comprising a prediction unit that predicts an intensity index related to the intensity of a mass spectrum of a compound to be predicted by inputting the physical index of the compound to be predicted and structural features of the compound to be predicted into a trained model that has been trained using a training database in which physical indexes related to atoms constituting the compound, calculated based on machine learning potential, are associated with intensity indexes related to the intensity of the mass spectrum of the compound.

6. The computer A learning method that performs a learning process to generate a trained model by learning using a learning database in which physical indicators related to the atoms that constitute a compound, calculated based on machine learning potential, are associated with intensity indicators related to the intensity of the mass spectrum of the compound.

7. On the computer, A learning program that executes a learning process to generate a trained model by learning using a learning database in which physical indicators related to the atoms that constitute a compound calculated based on machine learning potential are associated with intensity indicators related to the intensity of the mass spectrum of the compound.

8. The computer A prediction method comprising: executing a prediction step of predicting the intensity index relating to the intensity of the mass spectrum of a compound to be predicted by inputting the physical index of the compound to be predicted and structural features of the compound to be predicted into a trained model trained using a training database in which physical indexes relating to the atoms constituting the compound, calculated based on machine learning potential, and intensity indexes relating to the intensity of the mass spectrum of the compound are associated with each other.

9. On the computer, A prediction program that executes a prediction step of predicting the intensity index related to the intensity of the mass spectrum of a compound to be predicted by inputting the physical index of the compound to be predicted and structural features of the compound to be predicted into a trained model that has been trained using a training database in which physical indexes related to the atoms that constitute the compound, calculated based on machine learning potential, and intensity indexes related to the intensity of the mass spectrum of the compound are associated with each other.

Images