[0001] The present invention relates to a method, a program, and a device which perform calculation of a feature quantity, screening of a compound, and creation of a three-dimensional structure of a compound and particularly relates to a technique for searching for a pharmaceutical candidate compound.
2. Description of the Related Art
[0002] In the related art, in the drug discovery research using a computer, pharmaceutical candidate compounds (hereinafter, referred to as “hits”) have been searched for by preparing a library having tens of thousands to one hundred thousand of existing compounds, providing the structural formulae of the compounds, and examining the binding force with respect to one target protein. For example, US9373059B predicts a binding force by providing the structural formula of a compound. Further, JP5946045B also describes that a compound having a desired binding force is searched for gradually by repeating generation of a structural formula and prediction of a binding force (trial and error).
[0003] Further, JP4564097B describes that a search is performed using a descriptor referred to as “compound fingerprint”. The “descriptor” indicates information obtained from the structural formula of a compound, and the “compound fingerprint” indicates information related to the presence or absence of various functional groups. Such a descriptor has a characteristic in that “in a case where the descriptors of compounds are similar, the skeletons of the compounds are similar”.
SUMMARY OF THE INVENTION
[0005] Recently, since highly required target proteins have been complicated and difficult, it is difficult to find hits simply by screening libraries. Meanwhile, the theoretical number of compounds is (the 60th power of 10) even limiting the number to low molecules with a molecular weight of 500 or less. The number thereof is further increased in a case of widening the range to middle molecules with a molecular weight of approximately 1000, and thus there is still a possibility of finding hits, considering that the number of compounds
2

synthesized since the dawn of history is approximately (the ninth power of 10). However, it is almost impossible to examine the binding force with respect to all the astronomical numbers of compounds by experiments or simulations. Even in a case of examination of the binding force with respect to some compounds, the efficiency is low only by repeating trial and error as described in US9373059B and JP5946045B. Further, in the case of a descriptor (feature quantity) in the related art such as the fingerprint described in JP4564097B, the feature quantities of compounds are not necessarily similar even in a case where the compounds exhibit the same drug efficacy. Further, since the feature quantities did not accurately show the chemical properties of the target structure, the efficiency of search using the feature quantities was low.
[0006] As described above, in the related art, feature quantities do not accurately show the chemical properties of the target structures, and thus the efficiency of screening using the feature quantity and creation of a three-dimensional structure is low.
[0007] The present invention has been made in consideration of the above-described circumstances, and an object thereof is to provide a feature quantity calculating method, a feature quantity calculating program, and a feature quantity calculating device which enable calculation of a feature quantity accurately showing chemical properties of a target structure. Further, another object of the present invention is to provide a screening method and a screening program which enable efficient screening of a pharmaceutical candidate compound using a feature quantity. Further, still another object of the present invention is to provide a compound creating method which enable efficient creation of a three-dimensional structure of a pharmaceutical candidate compound using a feature quantity.
[0008] In order to achieve the above-described object, according to a first aspect of the present invention, there is provided a feature quantity calculating method comprising: a target structure designating step of designating a target structure formed of a plurality of unit structures having chemical properties; and a feature quantity calculating step of calculating a feature quantity obtained by quantifying, in a three-dimensional space, a degree of accumulation of one or more kinds of probes in a periphery of a three-dimensional structure of the target structure and calculating the feature quantity from the target structure using a generator formed through machine learning, in which the probe is a structure in which a plurality of points having a real electric charge and generating a van der Waals force are disposed to be separated from each other. [0009] Since the chemical properties of target structures are exhibited as the result of an
3

interaction between the target structure and one or more kinds of probes in the periphery thereof, the fact that target structures have a similarity in the degree of accumulation of probes indicates that the chemical properties of the target structures are similar. That is, target structures having similar feature quantities calculated according to the first aspect exhibit similar chemical properties. Therefore, according to the first aspect, the feature quantity accurately showing the chemical properties of a target structure can be calculated.
[0010] In the first aspect, the “generator” (prediction model, learned model) used for calculating a feature quantity can be constructed by “a method of constructing a generator, including a learning processing step of inputting [a data set for learning using information (three-dimensional images and the like) related to the three-dimensional structures of compounds as teacher data and the feature quantities as explanatory variables for a plurality of compounds] to the generator to perform learning through machine learning”. The generator may be constructed using a hierarchical neural network.
[0011] In the first aspect, the “point” constituting a probe is different from the mathematical “point” and may have a size (an actual atom, an actual ion, or the like). Further, a virtual point electric charge (one aspect of the “point”), a dipole, or the like may be used as a probe instead of an actual atom or the like. The probe can be selected according to a target compound (target structure).
[0012] Further, the feature quantity calculating method according to the first aspect can be performed by a computer realized using one or more of various processors such as a central processing unit (CPU) or a feature quantity calculating device, a feature quantity calculating system, or the like comprising such a computer. The same applies to the feature quantity calculating method, the screening method, and the compound creating method according to each of the following aspects.
[0013] In the feature quantity calculating method according to a second aspect, in the first aspect, a compound is designated as the target structure in the target structure designating step, and a first feature quantity which is a feature quantity obtained by quantifying, in the three-dimensional space, a degree of accumulation of amino acids as the probes in the periphery of the three-dimensional structure of the compound is calculated using the generator in the feature quantity calculating step. In the second aspect, the “probe”, the “target structure”, and the “plurality of unit structures” in the first aspect are respectively an amino acid, a compound, and a plurality of atoms. Further, the number of kinds of amino acids used for quantifying the degree of accumulation is not limited to one, and a peptide in which two or
4

more kinds of amino acids are bound may be used.
[0014] Similar to the first aspect, since the drug efficacy of a compound (the binding force with respect to the target protein) is locally exhibited as the result of an interaction between the compound and each amino acid (one aspect of a probe), in a case where the degree of accumulation of amino acids is similar between compounds, the compounds have similar binding forces (with respect to the target proteins). That is, compounds having similar feature quantities (first feature quantities) according to the second aspect exhibit similar drug efficacies. Therefore, according to the second aspect, the feature quantity accurately showing the chemical properties of a compound can be calculated. Further, in the second aspect, a compound such as a biological ligand which has a three-dimensional structure and whose binding to a target protein is known can be designated as a target structure.
[0015] In the feature quantity calculating method according to a third aspect, in the first aspect, a pocket structure bound to a pocket that is an active site of a target protein is designated as the target structure in the target structure designating step, and a second feature quantity which is a feature quantity obtained by quantifying, in the three-dimensional space, a degree of accumulation of amino acids as the probes in the periphery of a three-dimensional structure of the pocket structure formed of a plurality of virtual spheres is calculated using the generator in the feature quantity calculating step. In the third aspect, the “probe”, the “target structure”, and the “unit structure” in the first aspect are respectively an amino acid, a pocket structure, and a plurality of virtual spheres. The “active site” of the target protein indicates a site where the activity of the target protein is promoted or suppressed by binding a pocket structure, and the “virtual sphere” can be considered to have chemical properties such as the van der Waals radius and the electric charge.
[0016] Similar to the description of the second aspect, since the pocket structures having similar feature quantities (second feature quantities) according to the third aspect exhibit similar chemical properties, the feature quantity accurately showing the chemical properties of the pocket structure can be calculated according to the third aspect. Further, the pocket structure corresponds to a compound that is bound to a pocket of the target protein. Further, in the third aspect, as the result of actual measurement on the three-dimensional structure of the target protein, simulation based on position information and the like of the pocket can be used for calculation of the second feature quantity. Further, the measuring techniques (for example, an X-ray crystal structure, a nuclear magnetic resonance (NMR) structure, and a cryo-transmission electron microscopy (TEM) structure) are not limited as long as the
5

three-dimensional structure of the target protein is a three-dimensional structure with a resolution that enables identification of each residue of an amino acid.
[0017] In the feature quantity calculating method according to a fourth aspect, in the first aspect, a compound is designated as the target structure in the target structure designating step, and a third feature quantity which is a feature quantity obtained by quantifying, in the three-dimensional space, a degree of accumulation of the probes in the periphery of the three-dimensional structure of the compound which is the degree of accumulation using one or more selected from one or more kinds of nucleic acid bases, one or more kinds of lipid molecules, one or more kinds of monosaccharide molecules, water, or one or more kinds of ions formed of a plurality of atoms, as the probes is calculated using the generator in the feature quantity calculating step. In the fourth aspect, the “probe”, the “target structure”, and the “plurality of unit structures” in the first aspect are respectively one or more kinds of nucleic acid bases (the kind, the number, and the combination thereof may be optional), a compound, and a plurality of atoms.
[0018] In the present invention, deoxyribonucleic acid (DNA), ribonucleic acid (RNA), cell membranes, and polysaccharides, which are biopolymers (compounds) other than proteins can be treated as the targets of drugs. In the fourth aspect, the method of calculating the feature quantities of these target compounds is defined, and the probe is not an amino acid but another substance (a building block of each target). Specifically, in a case where the targets are DNA, RNA, cell membranes, and polysaccharides, the probes are respectively set as one or more kinds of nucleic acid bases, one or more kinds of nucleic acid bases, one or more kinds of lipid molecules, and one or more kinds of monosaccharide molecules. Further, water and one or more kinds of ions formed of a plurality of atoms may be considered in a case of the quantification of the degree of accumulation using these as probes. In the same manner as in the first to third aspects described above, since the drug efficacy of a compound (the binding force with respect to the target such as DNA) is locally exhibited as the result of an interaction between a compound and a nucleic acid base or the like (probe), in a case where the degree of accumulation of nucleic acid bases and the like is similar between compounds, the compounds have similar binding forces with respect to the targets. That is, compounds having similar feature quantities (third feature quantities) according to the fourth aspect exhibit similar drug efficacies. Therefore, according to the fourth aspect, the feature quantity accurately showing the chemical properties of a compound can be calculated. [0019] In the first to fourth aspects, an invariant conversion step of converting the feature
6

quantity into an invariant with respect to rotation and translation of the compound may be provided. By converting the feature quantity into an invariant, the feature quantity is easily handled and the data capacity can be reduced. The conversion of the feature quantity into an invariant can be performed by Fourier transform, angular integration of a correlation function, or the like.
[0020] In order to achieve the above-described object, according to a fifth aspect of the present invention, there is provided a feature quantity calculating program which causes a computer to execute the feature quantity calculating method according to any one of the first to fourth aspects. The “computer” in the fifth aspect can be realized using one or more of various processors such as a central processing unit (CPU). This “computer” may be a computer comprised in a feature quantity calculating device or a feature quantity calculating system. Further, a non-transitory recording medium on which a computer-readable code of the feature quantity calculating program according to the fifth aspect is recorded can also be exemplified as an aspect of the present invention.
[0021] In order to achieve the above-described object, according to a sixth aspect of the present invention, there is provided a feature quantity calculating device comprising: a target structure designation unit which designates a target structure formed of a plurality of unit structures having chemical properties; and a feature quantity calculation unit which calculates a feature quantity obtained by quantifying, in a three-dimensional space, a degree of accumulation of one or more kinds of probes in a periphery of a three-dimensional structure of the target structure and calculating the feature quantity from the target structure using a generator formed through machine learning, in which the probe is a structure in which a plurality of points having a real electric charge and generating a van der Waals force are disposed to be separated from each other. According to the sixth aspect, similarly to the first aspect, the feature quantity accurately showing the chemical properties of the target structure can be calculated. Further, the feature quantity calculating device according to the sixth aspect may perform the feature quantity calculating method according to present invention by reading the code of the feature quantity calculating program according to the present invention using various processors such as a CPU.
[0022] In order to achieve the above-described object, according to a seventh aspect of the present invention, there is provided a screening method of extracting a target compound which is bound to a target protein from a plurality of compounds, the method comprising: a storing step of storing a three-dimensional structure of a compound formed of a plurality of atoms and
7

the first feature quantity of the three-dimensional structure of the compound which is calculated using the feature quantity calculating method according to the second aspect in association with each other for each of the plurality of compounds; a feature quantity calculating step of calculating the first feature quantity of a ligand that is a compound whose binding to the target protein has been confirmed using the feature quantity calculating method according to the second aspect; a similarity calculating step of calculating a similarity between the first feature quantity of the plurality of compounds and the first feature quantity of the ligand; and a compound extracting step of extracting the target compound from the plurality of compounds based on the similarity. As the description for the second aspect, in a case where the ligand and the target compound have similar first feature quantities (the feature quantities three-dimensionally quantified), the drug efficacies of both the ligand and the target compound are similar. Therefore, according to the seventh aspect, a target compound having drug efficacy similar to that of the ligand can be extracted based on the first feature quantity so that screening of a pharmaceutical candidate compound can be efficiently performed. Further, a compound that has a similarity satisfying the target value (for example, greater than or equal to the threshold) may be extracted or a compound may be extracted in a descending order of the similarity in the compound extracting step.
[0023] In order to achieve the above-described object, according to an eighth aspect of the present invention, there is provided a screening method of extracting a target compound which is bound to a target protein from a plurality of compounds, the method comprising: a storing step of storing a three-dimensional structure of a compound formed of a plurality of atoms and the first feature quantity calculated using the feature quantity calculating method according to the second aspect in association with each other for each of the plurality of compounds; a feature quantity calculating step of calculating the second feature quantity of the pocket structure of the target protein using the feature quantity calculating method according to the third aspect; a similarity calculating step of calculating a similarity between the first feature quantity of the plurality of compounds and the second feature quantity of the pocket structure; and a compound extracting step of extracting the target compound from the plurality of compounds based on the similarity.
[0024] As the description for the third aspect, in a case where the pocket structure and the target compound have similar second feature quantities, the chemical properties of both the pocket structure and the target compound are similar. Therefore, according to the eighth aspect, a target compound having chemical properties similar to those of the pocket structure
8

is extracted so that screening of a pharmaceutical candidate compound can be efficiently performed. Since the pocket structure corresponds to the compound that is bound to the target protein, the feature quantity (the second feature quantity) of the pocket structure can be compared with the feature quantity (the first feature quantity) of the compound, and the similarity can be calculated.
[0025] In order to achieve the above-described object, according to a ninth aspect of the present invention, there is provided a screening method of extracting a target compound which is bound to a target biopolymer other than a protein from a plurality of compounds, the method comprising: a storing step of storing a three-dimensional structure of a compound formed of a plurality of atoms and the third feature quantity of the three-dimensional structure of the compound which is calculated using the feature quantity calculating method according to the fourth aspect in association with each other for each of the plurality of compounds; a feature quantity calculating step of calculating the third feature quantity of a binding compound that is a compound whose binding to the target biopolymer other than the protein has been confirmed using the feature quantity calculating method according to the fourth aspect; a similarity calculating step of calculating a similarity between the third feature quantity of the plurality of compounds and the third feature quantity of the binding compound; and a compound extracting step of extracting the target compound from the plurality of compounds based on the similarity. As described above for the fourth aspect, according to the present invention, DNA or the like which is a target biopolymer other than a protein can be handled, and in a case where the target compound and the binding compound that is bound to the target biopolymer have similar third feature quantities, the drug efficacies of both the target compound and the binding compound are similar. Therefore, according to the ninth aspect, a target compound having drug efficacy similar to that of the binding compound is extracted based on the third feature quantity so that screening of a pharmaceutical candidate compound can be efficiently performed.
[0026] In order to achieve the above-described object, according to a tenth aspect of the present invention, there is provided a screening method comprising: an input step of inputting information related to three-dimensional structures of a plurality of compounds; a feature quantity calculating step of calculating the feature quantity of each of the plurality of compounds using the feature quantity calculating method according to any one of first to fourth aspects based on the information; a characteristic value calculating step of calculating a physical property value indicating a physical property and/or a toxicity value indicating
9

toxicity of each of the plurality of compounds based on the information; and an extracting step of extracting compounds in which a similarity between the feature quantities satisfies a target value and the physical property value and/or the toxicity value satisfies a target value, from the plurality of compounds.
[0027] According to the feature quantities of the present invention (the feature quantities calculated by the feature quantity calculating method according to the first to fourth aspects), even in a case where an existing compound has a problem in terms of physical properties and toxicity, the skeleton of the compound can be significantly changed by recreating a novel compound having the same binding force to the target protein or the like (that is, the feature quantities of the novel compound and the existing compound are the same as each other), and thus the possibility of overcoming the problem of the physical properties and the toxicity can be increased. Further, in order to achieve both “maintenance of the binding force” and “avoidance of the problems of physical properties and toxicity”, it is effective to combine the feature quantities according to the present invention with indices showing the physical properties and the toxicity. The tenth aspect defines one aspect of the screening method based on such an idea, and compounds in which a similarity between the feature quantities satisfies the target value and the physical property value and/or the toxicity value satisfies the target value are extracted from the plurality of compounds using the feature quantity calculating method according to any one of the first to fourth aspects. Therefore, the screening of the pharmaceutical candidate compound can be efficiently performed. [0028] In the tenth aspect, it is preferable the feature quantities of compounds for which the similarity is determined are calculated according to the same aspect (using the same feature quantity calculating method). The same applies to the eleventh to thirteenth aspects described below.
[0029] In order to achieve the above-described object, according to an eleventh aspect of the present invention, there is provided a screening method comprising: an input step of inputting information related to three-dimensional structures of a plurality of compounds; a first feature quantity calculating step of calculating the feature quantity of each of the plurality of compounds using the feature quantity calculating method according to any one of the first to fourth aspects based on the information; a second feature quantity calculating step of calculating a structural feature quantity indicating a compound structure of each of the plurality of compounds based on the information; and an extracting step of extracting compounds in which a similarity between the feature quantities satisfies a target value and a
10

similarity between the structural feature quantities is less than or equal to a threshold, from the plurality of compounds.
[0030] In order to ensure that a novel compound obtained by search has a skeleton different from the skeleton of the original (existing) compound serving as the starting point of the search, it is effective to impose a condition that “the feature quantities (according to the present invention) of the original compound and the novel compound are the same as each other, but the structural formulae thereof are different from each other” in a case of the search of the compound. The eleventh aspect defines one aspect of the screening method based on such an idea, and compounds in which a similarity between the feature quantities satisfies the target value and a similarity between the structural feature quantities is less than or equal to the threshold are extracted from the plurality of compounds using the feature quantity calculating method according to any one of the first to fourth aspects. Therefore, the screening of the pharmaceutical candidate compound can be efficiently performed. Further, it is effective to use the index of the similarity between the structural formulae to indicate that “the structural formulae thereof are different from each other”, but many known indices which have been suggested can be used as such an index.
[0031] In order to achieve the above-described object, according to a twelfth aspect of the present invention, there is provided a screening method comprising: an input step of inputting information related to three-dimensional structures of a plurality of compounds; a feature quantity calculating step of calculating the feature quantity of each of the plurality of compounds using the feature quantity calculating method according to any one of the first to fourth aspects based on the information; a predicting step of predicting a binding force using a prediction model that outputs the binding force between each of the plurality of compounds and another compound based on the feature quantity; and an extracting step of extracting compounds, in which the binding force satisfies a target value, from the plurality of compounds. In the screening method according to the twelfth aspect, compounds in which the binding force satisfies the target value can be efficiently extracted (screening of the pharmaceutical candidate compound can be efficiently performed) by changing the information serving as the starting point of the search to information related to “the prediction model of the binding force based on the feature quantity of the present invention” and “the target value of the binding force (which can be set to be higher than the current value)” from the information related to only the structural formula of one compound. [0032] In order to achieve the above-described object, according to a thirteenth aspect of the
11

present invention, there is provided a screening method comprising: an input step of inputting information related to three-dimensional structures of a plurality of compounds; a setting step of setting a target value of a binding force between each of the plurality of compounds and another compound; an estimating step of estimating the feature quantity corresponding to the target value using a prediction model that outputs the binding force based on the feature quantity calculated by the feature quantity calculating method according to any one of the first to fourth aspects; a feature quantity calculating step of calculating the feature quantity of each of the plurality of compounds using the feature quantity calculating method according to any one of the first to fourth aspects based on the information; and an extracting step of extracting compounds in which a similarity between the calculated feature quantity and the estimated feature quantity is greater than or equal to a threshold, from the plurality of compounds based on the calculation results. In the twelfth aspect described above, the compound satisfying the “target value of the binding force” is extracted by predicting the binding force using the “prediction model”. Meanwhile, in the screening method according to the thirteenth aspect, the “prediction model” is used in the opposite manner to that of the twelfth aspect. Specifically, the feature quantity (the feature quantity according to any one of the first to fourth aspects) is estimated from the “target value of the binding force” using the “prediction model” to search a compound having the estimated feature quantity. Even in a case of using such a method, screening of the pharmaceutical candidate compound can be efficiently performed as in the twelfth aspect.
[0033] Further, in the twelfth and thirteenth aspects, the “prediction model of the binding force” can be generated from, for example, a set of structural formulae of a plurality of binding compounds and experimental values of the binding force thereof.
[0034] In order to achieve the above-described object, according to a fourteenth aspect of the present invention, there is provided a screening program which causes a computer to execute the screening method according to any one of the seventh to thirteenth aspects. The “computer” in the fourteenth aspect can be realized using one or more of various processors such as a CPU. This “computer” may be a computer comprised in a screening device or a screening system. Further, a non-transitory recording medium on which a computer-readable code of the screening program according to the fourteenth aspect is recorded can also be exemplified as an aspect of the present invention.
[0035] In order to achieve the object described above, according to a fifteenth aspect of the present invention, there is provided a compound creating method of creating a
12

three-dimensional structure of a target compound that is bound to a target protein from a plurality of compounds, the method comprising: a storing step of storing a three-dimensional structure of a compound formed of a plurality of atoms and the first feature quantity calculated using the feature quantity calculating method according to the second aspect in association with each other for each of the plurality of compounds; a feature quantity calculating step of calculating the first feature quantity of a ligand that is a compound whose binding to the target protein has been confirmed using the feature quantity calculating method according to the second aspect; a generator constructing step of constructing a generator through machine learning using the three-dimensional structures of the plurality of compounds as teacher data and the first feature quantity as an explanatory variable; and a compound three-dimensional structure generating step of generating the three-dimensional structure of the target compound from the first feature quantity of the ligand using the generator.
[0036] In the screening methods according to the seventh to thirteenth aspects described above, a compound that is compatible with a ligand or a target protein is found among a plurality of compounds whose structural formulae have already been determined (written down). Accordingly, after the feature quantity of the compound is calculated, a method of extracting the compound based on the similarity with the feature quantity of the separately calculated ligand or the pocket structure of the target protein, that is, a search method is employed. Therefore, in a case where the correspondence between the structural formula of the compound and the feature quantity thereof is recorded, a structural formula having a high similarity (greater than or equal to the threshold) can be found. Meanwhile, in the fifteenth aspect, a structural formula of a compound having a feature quantity similar to the feature quantity (the first feature quantity) of the ligand (accordingly, the drug efficacies are similar) is generated without performing search.
[0037] The generation of the structural formula in a case where the feature quantity has been provided can be performed using a generator constructed through machine learning. Specifically, in the fifteenth aspect, a generator is constructed through machine learning (the learning method is not particularly limited) using the three-dimensional structures of the compounds as teacher data and the first feature quantity as an explanatory variable, and a three-dimensional structure of the target compound is generated from the first feature quantity of the ligand using the generator. In the fifteenth aspect, since search is not performed, the three-dimensional structure of the compound can be generated even in a case of “no solution was found as the result of screening search”, and thus the three-dimensional structure of the
13

pharmaceutical candidate compound can be efficiently created. Further, the
three-dimensional structure to be generated in the fifteenth aspect is affected by the features of
the compound provided as teacher data. Therefore, by selecting the features of the
compound to be provided as teacher data, a compound having a three-dimensional structure
with different features can be generated. For example, a compound having a
three-dimensional structure that is easily synthesized can be generated by providing a compound that is easily synthesized as teacher data.
[0038] In order to achieve the object described above, according to a sixteenth aspect of the present invention, there is provided a compound creating method of creating a three-dimensional structure of a target compound that is bound to a target protein from a plurality of compounds, the method comprising: a storing step of storing a three-dimensional structure of a compound formed of a plurality of atoms and the first feature quantity calculated using the feature quantity calculating method according to the second aspect in association with each other for each of the plurality of compounds; a feature quantity calculating step of calculating the second feature quantity of the pocket structure of the target protein using the feature quantity calculating method according to the third aspect; a generator constructing step of constructing a generator through machine learning using the three-dimensional structures of the plurality of compounds as teacher data and the first feature quantity as an explanatory variable; and a compound three-dimensional structure generating step of generating the three-dimensional structure of the target compound from the second feature quantity of the pocket structure using the generator. According to the sixteenth aspect, similar to the fifteenth aspect, a structural formula of a compound having a feature quantity similar to the feature quantity (the second feature quantity) of the pocket structure (accordingly, the drug efficacies are similar) is generated without performing search, and thus the three-dimensional structure of the pharmaceutical candidate compound can be efficiently created. Further, similar to the fifteenth aspect, by selecting the features of the compound to be provided as teacher data, a compound having a three-dimensional structure with different features can be generated.
[0039] In order to achieve the above-described object, according to a seventeenth aspect of the present invention, there is provided a compound creating method of creating a three-dimensional structure of a target compound that is bound to a target biopolymer other than a protein from a plurality of compounds, the method comprising: a storing step of storing a three-dimensional structure of a compound formed of a plurality of atoms and a third feature
14

quantity in association with each other for each of the plurality of compounds; a feature quantity calculating step of calculating the third feature quantity of a binding compound that is a compound whose binding to the target biopolymer other than the protein has been confirmed using the feature quantity calculating method according to the fourth aspect; a generator constructing step of constructing a generator through machine learning using the three-dimensional structures of the plurality of compounds as teacher data and the third feature quantity as an explanatory variable; and a compound three-dimensional structure generating step of generating the three-dimensional structure of the target compound from the third feature quantity of the binding compound using the generator.
[0040] According to the seventeenth aspect, similar to the fifteenth and sixteenth aspects, a structural formula of a compound having a feature quantity similar to the feature quantity (the third feature quantity) of the binding compound (accordingly, the drug efficacies are similar) is generated without performing search, and thus the three-dimensional structure of the pharmaceutical candidate compound can be efficiently created. Further, similar to the fifteenth and sixteenth aspects, by selecting the features of the compound to be provided as teacher data, a compound having a three-dimensional structure with different features can be generated.
[0041] As described above, according to the feature quantity calculating method, the feature quantity calculating program, and the feature quantity calculating device of the present invention, it is possible to calculate a feature quantity that accurately shows chemical properties of a target structure. Further, according to the screening method and the screening program of the present invention, screening of a pharmaceutical candidate compound can be efficiently performed. Further, according to the compound creating method of the present invention, the three-dimensional structure of a pharmaceutical candidate compound can be efficiently created.
BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Fig. 1 is a block diagram showing a configuration of a screening device according to a first embodiment.
Fig. 2 is a block diagram showing a configuration of a processing unit.
Fig. 3 is a diagram showing information stored in a storage unit.
Fig. 4 is a diagram showing a state in which structure information of a compound and a feature quantity thereof are stored in association with each other.
Fig. 5 is a flowchart showing a procedure for calculating a three-dimensional AAM
15

feature quantity of a compound.
Figs. 6A and 6B are diagrams showing an example of three-dimensionalization of a structural formula.
Figs. 7A and 7B are diagrams showing an example of a three-dimensional AAM feature quantity.
Fig. 8 is a table showing an example of a three-dimensional AAM feature quantity.
Fig. 9 is a flowchart showing a procedure for a three-dimensional AAM descriptor for a pocket structure.
Figs. 10A to 10C are conceptual views showing a state of a three-dimensional AAM descriptor for a pocket structure.
Figs. 11A and 11B are graphs showing an example of an invariant AAM descriptor.
Figs. 12A and 12B show an example of compounds with similar invariant AAM descriptors.
Fig. 13 is a graph showing easiness of finding a hit in a case where an invariant AAM descriptor is used.
Fig. 14 is a graph showing the time of search for a hit in a case where an invariant AAM descriptor is used.
Fig. 15 is another graph showing the time of search for a hit in a case where an invariant AAM descriptor is used.
Fig. 16 is a flowchart showing a procedure for extracting a target compound based on the similarity between descriptors.
Figs. 17A and 17B are tables showing an example of a result of extracting a target compound based on the similarity between descriptors.
Fig. 18 is another flowchart showing a procedure for extracting a target compound based on the similarity between descriptors.
Figs. 19A and 19B are other tables showing an example of a result of extracting a target compound based on the similarity between the descriptors.
Fig. 20 is a block diagram showing a configuration of a compound creating device according to a second embodiment.
Fig. 21 is a diagram showing a configuration of a processing unit.
Fig. 22 is a diagram showing information stored in a storage unit.
Fig. 23 is a flowchart showing a procedure for generating a three-dimensional structure in the case where a ligand is input.
16

Fig. 24 is a diagram showing a state of generating a three-dimensional structure using a result of machine learning.
Fig. 25 is a graph showing a relationship between the number of interlayers and the cos similarity.
Figs. 26A and 26B are diagrams showing an example of generating a three-dimensional structure.
Fig. 27 is a flowchart showing a procedure for generating a three-dimensional structure in a case where a target protein is input.
Fig. 28 is a block diagram showing a configuration of a pharmaceutical candidate compound search device according to a third embodiment.
Fig. 29 is a diagram showing a configuration of a processing unit.
Fig. 30 is a diagram showing information stored in a storage unit.
Fig. 31 is a diagram showing a comparison result of easiness of finding a hit.
Figs. 32A and 32B are views showing a configuration example of a hierarchical network.
Fig. 33 is a flowchart showing a screening process in a case where the feature quantity according to the present invention and another feature quantity are used in combination.
Fig. 34 is a diagram showing another example of a configuration of a processing unit.
Figs. 35A and 35B are diagrams showing a state of screening in a case where the feature quantity according to the present invention and a physical property value are used in combination.
Fig. 36 is another diagram showing a state of screening in a case where the feature quantity according to the present invention and a physical property value are used in combination.
Fig. 37 is still another diagram showing a state of screening in a case where the feature quantity according to the present invention and a physical property value are used in combination.
Fig. 38 is a flowchart showing a screening process in a case where the feature quantity according to the present invention and another feature quantity are used in combination.
Fig. 39 is a diagram showing still another example of a configuration of a processing unit.
17

Figs. 40A and 40B are diagrams showing a state of screening in a case where the feature quantity according to the present invention and a structural feature quantity are used in combination.
Fig. 41 is another diagram showing a state of screening in a case where the feature quantity according to the present invention and a structural feature quantity are used in combination.
Fig. 42 is still another diagram showing a state of screening in a case where the feature quantity according to the present invention and a structural feature quantity are used in combination.
Figs. 43A and 43B are diagrams showing an output example in a case where the feature quantity according to the present invention and another feature quantity are used in combination.
Figs. 44A and 44B are other diagrams showing an output example in a case where the feature quantity according to the present invention and another feature quantity are used in combination.
Fig. 45 is another flowchart showing a screening process in a case where the feature quantity according to the present invention and a prediction model are used.
Fig. 46 is a diagram showing still another example of a configuration of a processing unit.
Fig. 47 is still another flowchart showing a screening process in a case where the feature quantity according to the present invention and a prediction model are used.
Fig. 48 is a flowchart showing a process of creating a compound in a case where the feature quantity according to the present invention and another feature quantity are used in combination.
Fig. 49 is another flowchart showing a process of creating a compound in a case where the feature quantity according to the present invention and another feature quantity are used in combination.
Fig. 50 is still another flowchart showing a process of creating a compound in a case where the feature quantity according to the present invention and another feature quantity are used in combination.
Fig. 51 is even still another flowchart showing a process of creating a compound in a case where the feature quantity according to the present invention and another feature quantity are used in combination.
18

Fig. 52 is a flowchart showing a process of creating a compound in a case where the feature quantity according to the present invention and a structural feature quantity are used in combination.
Fig. 53 is another flowchart showing a process of creating a compound in a case where the feature quantity according to the present invention and a structural feature quantity are used in combination.
Fig. 54 is still another flowchart showing a process of creating a compound in a case where the feature quantity according to the present invention and a structural feature quantity are used in combination.
Fig. 55 is even still another flowchart showing a process of creating a compound in a case where the feature quantity according to the present invention and a structural feature quantity are used in combination.
Fig. 56 is a flowchart showing a process of creating a compound in a case where the feature quantity according to the present invention and a prediction model are used.
Fig. 57 is another flowchart showing a process of creating a compound in a case where the feature quantity according to the present invention and a prediction model are used.
Fig. 58 is still another flowchart showing a process of creating a compound in a case where the feature quantity according to the present invention and a prediction model are used.
Fig. 59 is a flowchart showing a screening process according to the present invention.
Fig. 60 is a flowchart showing a process of creating a compound according to the present invention.
Fig. 61 is another flowchart showing a screening process according to the present invention.
Fig. 62 is another flowchart showing a process of creating a compound according to the present invention.
Fig. 63 is still another flowchart showing a screening process according to the present invention.
Fig. 64 is still another flowchart showing a process of creating a compound according to the present invention.
Fig. 65 is still another flowchart showing compound information processing according to the present invention.
Fig. 66 is a diagram showing an output example of the degree of accumulation of amino acids.
19

Fig. 67 is another diagram showing an output example of the degree of accumulation of amino acids.
Fig. 68 is a diagram showing an example of a two-dimensional display of the degree of accumulation of amino acids.
Figs. 69A and 69B are other diagrams showing an example of a two-dimensional display of the degree of accumulation of amino acids.
Fig. 70 is still another diagram showing an example of a two-dimensional display of the degree of accumulation of amino acids.
Fig. 71 is a diagram showing an example of a graph display of the degree of accumulation of amino acids.
Fig. 72 is another diagram showing an example of a graph display of the degree of accumulation of amino acids.
Figs. 73A and 73B are diagrams showing an example of a pseudo three-dimensional display of a feature quantity.
Fig. 74 is a diagram showing the state in which a difference in the degree of accumulation of amino acids is expressed as the similarity.
Fig. 75 is a graph showing an output example of the performance of a prediction model for the binding force.
Fig. 76 is a diagram showing a configuration example of a compound search system according to the present invention.
Fig. 77 is a diagram showing an example of a functional configuration of a compound search server.
Fig. 78 is a diagram showing an example of a functional configuration of a client device.
Fig. 79 is a diagram showing an example of a functional configuration of a compound creation server.
Fig. 80 is a diagram showing another example of a functional configuration of a compound search server.
DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0043] Hereinafter, embodiments of a feature quantity calculating method, a feature quantity calculating program, a feature quantity calculating device, a screening method, a screening program, and a compound creating device of the present invention will be described in detail with reference to the accompanying drawings.
20

[0044]
Fig. 1 is a block diagram showing a configuration of a screening device 10 (a feature quantity calculating device or a screening device) according to a first embodiment. The screening device 10 is a device that calculates a feature quantity of a compound (target structure) and/or a pocket structure (target structure) and extracts (screens) a target compound, and can be realized using a computer. As shown in Fig. 1, the screening device 10 includes a processing unit 100, a storage unit 200, a display unit 300, and an operation unit 400, and these units are connected to one another to transmit and receive necessary information. These constituent elements may be installed by employing various installation forms. Respective constituent elements may be installed in one site (in one housing, one room, or the like) or may be installed in places separated from each other and connected via a network. Further, the screening device 10 is connected to an external server 500 and an external database 510 such as a Protein Data Bank (PDB) via a network NW such as the Internet, and information related to structural formulae of compounds and crystal structures of proteins can be obtained as necessary. [0045]
Fig. 2 is a diagram showing a configuration of the processing unit 100. The processing unit 100 includes an information input unit 110, a feature quantity calculation unit 120, a similarity calculation unit 130, a compound extraction unit 140, a display control unit 150, a CPU 160 (CPU: Central Processing Unit), a ROM 170 (ROM: Read Only Memory), and a RAM 180 (RAM: Random Access Memory).
[0046] The information input unit 110 inputs information related to a structural formula of a compound, an X crystal structure of a target protein, and a pocket position via a recording medium interface such as a DVD drive (not shown) or a semiconductor memory terminal and/or a network NW. The feature quantity calculation unit 120 (the target structure designation unit, the three-dimensional structure generation unit, the feature quantity calculation unit, and the invariant conversion unit) calculates the feature quantity according to the present invention. The similarity calculation unit 130 (the similarity calculation unit) calculates the similarity between the calculated feature quantities. The compound extraction unit 140 (the compound extraction unit) extracts a target compound from a plurality of compounds based on the similarity. The display control unit 150 controls input information and display of the process result on the monitor 310. The process of calculation of the feature quantity and screening of the target compound using these functions of the processing
21

unit 100 will be described below in detail. Further, the process using these functions is performed under the control of the CPU 160.
[0047] The function of each unit of the processing unit 100 described above can be realized using various processors. Various processors include a CPU that is a general-purpose processor that executes software (program) to realize various functions. Further, the various processors described above include a graphics processing unit (GPU) serving as a processor specialized in image processing and a programmable logic device (PLD) serving as a processor that can change the circuit configuration after manufacture of a field programmable gate array (FPGA). Further, the above-described various processors include an exclusive electric circuit serving as a processor having a circuit configuration designed exclusively for executing a specific process such as an application specific integrated circuit (ASIC). [0048] The functions of each unit may be realized by one processor, or may be realized by a plurality of same or different processors (for example, a plurality of FPGAs, a combination of a CPU and an FPGA, or a combination of a CPU and a GPU). Further, a plurality of functions may be realized by one processor. As an example in which a plurality of functions are configured by one processor, first, as represented by a computer such as a client or a server, a form of one processor which is configured by a combination of one or more CPUs and software and can be realized as a plurality of functions is exemplified. Second, as represented by a system-on-chip (SoC) or the like, there is a form in which a processor that realizes the functions of the entire system by one integrated circuit (IC) chip is used. As described above, various functions are configured using one or more of the above-described various processors as a hardware structure. Further, the hardware structure of these various processors is more specifically an electric circuit (circuitry) in which circuit elements such as semiconductor elements are combined.
[0049] In a case where the above-described processor or electric circuit executes software (program), a code that can be read by the computer (for example, various processors or electric circuits constituting the processing unit 100, and/or a combination thereof) to execute the software is stored in a non-transitory recording medium such as a ROM 170 (see Fig. 2), and the processor refers to the software. The software stored in the non-transitory recording medium includes the feature quantity calculating method and a program (a feature quantity calculating program and a screening program) for executing a target compound extraction process according to the present invention. The code may be recorded on non-transitory recording media such as various magneto-optical recording devices and semiconductor
22

memories instead of the ROM 170. During the process using software, for example, the RAM 180 is used as a temporary storage area, and the data stored in, for example, an electronically erasable and programmable read only memory (EEPROM) (not shown) can also be referred to. [0050]
The storage unit 200 is formed of non-transitory recording media such as a digital
versatile disk (DVD), a hard disk, and various semiconductor memories and control units
thereof and stores images and information shown in Fig. 3. Structure information 210
includes the structural formula of the compound, the three-dimensional structure of the target
protein, and the position of the pocket. Three-dimensional structure information 220 is
information related to the three-dimensional structure of the compound and/or the pocket
structure generated from the structure information 210. A three-dimensional AAM descriptor
230 is a feature quantity obtained by quantifying, in a three-dimensional space, the degree of
accumulation of one or more kinds of amino acids in the periphery of the three-dimensional
structure of the compound or the pocket structure, and is calculated according to the feature
quantity calculating method described below. Further, “AAM” stands for “Amino Acid
Mapping”. An invariant AAM descriptor 240 is a feature quantity obtained by converting the
three-dimensional AAM descriptor 230 into an invariant with respect to rotation and
translation of the compound or the pocket structure. Similarity information 250 is
information related to the similarity between the feature quantities, and a compound extraction result 260 is information related to the target compound extracted based on the similarity. [0051] Fig. 4 shows a state in which the structure information 210, the three-dimensional structure information 220, the three-dimensional AAM descriptor 230, and the invariant AAM descriptor 240 are associated with one another and stored in the storage unit 200 for N compounds (N represents an integer of 2 or greater). In Fig. 4, for example, a structural formula can be used as the structure information 210, and a three-dimensionalized structural formula (described below) can be used as the three-dimensional structure information 220. In Fig. 4, the three-dimensional AAM descriptor 230 (described as “ga(r)”; a is a subscript indicating the kind of amino acid) and the invariant AAM descriptor 240 (described as “Fab (s)”; a and b are subscripts indicating the kinds of amino acids) corresponding to the three-dimensional AAM descriptor 230 for each of twenty kinds of amino acids for each compound are stored in association with each other. The three-dimensional AAM descriptor 230 and the invariant AAM descriptor 240 may be stored not for all twenty kinds of amino
23

acids but for some amino acids according to the number of descriptors used for screening. [0052] A plurality of sets (libraries) of information as shown in Fig. 4 may be stored in the storage unit 200. Fig. 4 shows a state in which information related to compounds is stored, and information related to target proteins can be stored with the same configuration. Further, a method for calculating a three-dimensional AAM descriptor and/or an invariant AAM descriptor using such structure information and three-dimensional structure information will be described below. [0053]
The display unit 300 includes the monitor 310 (display device) and can display input images, images and information stored in the storage unit 200, results of the process carried out by the processing unit 100, and the like. The operation unit 400 includes a keyboard 410 and a mouse 420 as an input device and/or a pointing device, and execution of the feature quantity calculating method according to the present invention and the operation required for extraction of the target compounds can be performed by a user through these devices and the screen of the monitor 310 (described later). The operations that can be performed by the user include, for example, a processing mode, the kind of descriptor to be calculated, a descriptor used for screening, and designation of a threshold with respect to the similarity. [0054]
In the screening device 10 with the above-described configuration, calculation of a feature quantity (descriptor) and/or extraction of a target compound can be performed according to the user’s instruction via the operation unit 400. Hereinafter, the details of each process will be described. [0055]
The screening device 10 is capable of calculating a three-dimensional AAM descriptor and/or an invariant AAM descriptor according to the user's instruction via the operation unit 400. [0056]
Fig. 5 is a flowchart showing a procedure for calculating a three-dimensional AAM descriptor for a compound (target structure). The ligand is a compound whose binding to the target protein has been confirmed, and the three-dimensional AAM descriptor can be calculated according to the procedure of Fig. 5. In Step S100, the information input unit 110 inputs a structural formula of a compound according to an operation by the user. In this manner, the compound represented by the input chemical formula is designated as the target
24

structure (target structure designating step).
[0057] The feature quantity calculation unit 120 three-dimensionalizes the input structural
formula to generate a three-dimensional structure of a compound formed of a plurality of
atoms (a plurality of unit structures having chemical properties) (Step S102: a
three-dimensional structure generating step). Various methods are known for
three-dimensionalization of a structural formula, and the present invention is not particularly limited to the method used in Step S102. Figs. 6A and 6B show an example of a three-dimensionalized structural formula. Fig. 6A shows an input structural formula and Fig. 6B shows a three-dimensionalized structural formula.
[0058] The feature quantity calculation unit 120 calculates a spatial distribution ΔGaμ(r) of free energy felt by each atom “μ” of an amino acid “a” (a represents a number representing the kind of amino acid; 1 to 20) (Step S104; feature quantity calculating step). As a method of calculating ΔGaμ(r), a molecular dynamics (MD) method can be employed, but the present invention is not limited thereto. The amino acid for calculating the feature quantity may be a predetermined kind of amino acid or may be determined according to the user's instruction (one or more kinds of amino acids may be used, and a plurality of kinds of amino acids may also be used).
[0059] The feature quantity calculation unit 120 calculates a distribution function gaμ(r) of each atom “μ” of the amino acid “a” from ΔGaμ(r) (Step S106: feature quantity calculating step). gaμ(r) is represented by Equation (1) in a case where T is set as room temperature and KB is set as a Boltzmann constant. [0060]

[0061] The feature quantity calculation unit 120 calculates a distribution function ga(r) of the center of gravity of an amino acid from the distribution function gaμ(r) (Step S108: feature quantity calculating step). For the calculation, gaμ(r) is geometrically averaged for each atom “μ”. This distribution function ga(r) is a three-dimensional AAM descriptor obtained by quantifying, in a three-dimensional space, the degree of accumulation of one or more kinds of amino acids “a” in the periphery of the three-dimensional structure of the compound. The feature quantity calculation unit 120 stores the calculated three-dimensional AAM descriptor in the storage unit 200 as the three-dimensional AAM descriptor 230 in association with the structure information (structure information 210) and the three-dimensional structure
25

information (three-dimensional structure information 220) of the compound. (See Fig. 4). [0062] Figs. 7A and 7B are an example of a three-dimensional AAM descriptor of the compound shown in Figs. 6A and 6B. Fig. 7A shows a three-dimensional AAM descriptor of alanine, and Fig. 7B shows a three-dimensional AAM descriptor of valine. In Figs. 7A and 7B, the dark region is a region where the degree of accumulation of amino acids (existence probability) is high. Fig. 8 is a table showing another example of the three-dimensional AAM descriptor of the compound shown in Figs. 6A and 6B, and the three-dimensional AAM descriptor is shown in a direction 1, a direction 2, and a direction 3 (the first, second, and third rows of the table, respectively) which are different from one another. The left columns of the table show the three-dimensional AAM descriptor (the level surface with respect to the threshold), and the right columns of the table show the three-dimensional AAM descriptor (the level surface with respect to the threshold) and the three-dimensional structure of the compound. [0063]
In the screening device 10, a pocket structure that is bound to a target protein instead of a compound is designated as a target structure, and the feature quantity (the three-dimensional AAM descriptor) of this pocket structure can be calculated. The pocket structure is a target structure that is bound to a pocket, which is an active site of the target protein, and the “active site” indicates a site where the activity of the target protein is promoted or suppressed by the binding of the pocket structure. Fig. 9 is a flowchart showing a procedure for calculating the three-dimensional AAM descriptor for the pocket structure. Figs. 10A to 10C are conceptual views showing a state of the three-dimensional AAM descriptor for the pocket structure.
[0064] In the flowchart of Fig. 9, the information input unit 110 inputs the actual measurement of the three-dimensional structure of the target protein and the position information of the pocket (Step S200: target structure designating step). Fig. 10A shows a pocket PO in a target protein TP. The pocket structure is designated as the target structure by the process of Step S200.
[0065] The feature quantity calculation unit 120 packs a plurality of virtual spheres (a plurality of unit structures having chemical properties) into the pocket of the target protein (Step S202: the target structure designating step or the three-dimensional structure generating step). The “virtual sphere” can be considered to have chemical properties such as a van der Waals radius and an electric charge, and “packing the virtual spheres” can be performed by simulation (for
26

example, a molecular dynamics method). A collection of the packed virtual spheres (three-dimensional structure) can be obtained as a three-dimensional structure of the pocket structure (target structure) in Step S202 (Step S204: three-dimensional structure generating step). Fig. 10B shows an example of a pocket structure PS of the target protein TP. [0066] The feature quantity calculation unit 120 three-dimensionally quantifies the degree of accumulation of one or more kinds of amino acids in the periphery of the pocket structure using actual measurement of the three-dimensional structure of the target protein (Step S206: feature quantity calculating step). Practically, it is possible to read out what kinds of amino acids are accumulated in the periphery of the pocket structure. Fig. 10C shows that three kinds of amino acids A1, A2, and A3 are accumulated in the periphery of the pocket structure PS. Further, the number of amino acids for quantifying the degree of accumulation may be one or more (a plurality of kinds of amino acids may be used). In addition, a predetermined kind of amino acid may be quantified or an amino acid which has been set according to the operation of the user may be quantified. The feature quantity calculation unit 120 stores the calculated three-dimensional AAM descriptor in the storage unit 200 as the three-dimensional AAM descriptor 230 in association with the structure information (structure information 210) and the three-dimensional structure information (three-dimensional structure information 220) of the compound (see Figs. 3 and 4; the storing step). In a case where the invariant AAM descriptor described below has been calculated, the feature quantity calculation unit 120 associates the three-dimensional AAM descriptor with the invariant AAM descriptor. [0067]
The above-described three-dimensional AAM descriptor indicates the degree of three-dimensional accumulation of amino acids. However, in a case where shift or rotation of the center of gravity occurs even though the compounds are the same as each other, the value changes and the data capacity is large because the information is three-dimensional information. Therefore, in the screening device 10 according to the first embodiment, “an invariant AAM descriptor obtained by converting the three-dimensional AAM descriptor into an invariant with respect to rotation and translation of the compound” (the invariant feature quantity) can be calculated in addition to or instead of the three-dimensional AAM descriptor. Further, conversion into an invariant can be performed according to the same procedures in both cases of a compound and a pocket structure. In a case where a three-dimensional AAM descriptor of a compound is used, an invariant AAM descriptor of the compound is obtained. Further, in a case where a three-dimensional AAM descriptor of the pocket structure is used,
27

an invariant AAM descriptor of the pocket structure is obtained.
[0068] The feature quantity calculation unit 120 (invariant conversion unit) calculates fa(k)
using Fourier transform as shown in Equation (2) (invariant conversion step). As described
above, “a” is a subscript (1 to 20) indicating the kind of amino acid. Further, “i” is an
imaginary unit.
[0069]

[0070] The feature quantity calculation unit 120 is capable of calculating Fab(s) (invariant feature quantity) that is an invariant AAM descriptor using the fa (k) according to Equation (3) (invariant conversion step). In Equation (3), an invariant AAM descriptor is calculated by angular integration of a correlation function using three-dimensional AAM descriptors (ga(r) and gb(r)) of two different kinds of amino acids (denoted by “a” and “b”). Further, the combination of the two kinds of amino acids used for calculation of the invariant AAM descriptors among twenty kinds of amino acids is not particularly limited. [0071]

[0072] In Equation (3), a delta function is used for conversion into an invariant. However, as shown in Equation (4), conversion into an invariant can be performed using an optional function (h(k2-s)). [0073]

[0074] Figs. 11A and 11B show an example of the invariant AAM descriptor calculated in the
above-described manner. Fig. 11A shows a real part of F12(s) (an invariant AAM descriptor
of amino acids 1 and 2), which is an invariant AAM descriptor, and Fig. 11B shows an
imaginary part. In this manner, the conversion into an invariant can be performed while
information related to the interaction between amino acids is maintained by performing
conversion into an invariant using a three-dimensional AAM descriptor of two different kinds
of amino acids so that compound comparison (drug efficacy determination) based on the
feature quantities (the invariant feature quantities) can be accurately performed.
[0075] The feature quantity calculation unit 120 stores the calculated invariant AAM
descriptor in the storage unit 200 as the invariant AAM descriptor 240 in association with the
28

structure information (structure information 210), the three-dimensional structure information (three-dimensional structure information 220), and the original three-dimensional AAM descriptor 230 of the compound (see Figs. 3 and 4; the storing step). In the first embodiment, since the invariant AAM descriptor is calculated using the three-dimensional AAM descriptor of two different kinds of amino acids, a plurality of associations between the three-dimensional AAM descriptor and the invariant AAM descriptor may be present. [0076]
The effectiveness of the invariant AAM descriptor calculated by the above-described process will be described. [0077]
Fig. 12A shows a structural formula of a ligand for a protein ABL1 (an example of a target protein). The binding force of this ligand is at a level of 1 μM with an IC50 (50% inhibitory concentration). Meanwhile, Fig. 12B shows a structural formula of a compound having almost the same invariant AAM descriptor as that of the ligand. In a case where the activity of this compound is actually measured, the activity is at the same level as that of the ligand. That is, Figs. 11A and 11B are an example showing that compounds having similar invariant AAM descriptors have similar drug efficacies. As described above, according to the first embodiment, a feature quantity (invariant AAM descriptor) accurately showing the chemical properties of the target structure can be obtained. Further, the index or unit indicating the strength of the binding force can be classified into two parts. One relates to the binding energy, and kJ/mol and kcal/mol are used as the units. The other one relates to IC50 or Kd (binding dissociation constant) described above, and the concentration M (= mol/L; molar) is used as the unit. [0078]
The easiness of finding hits using invariant AAM descriptors is evaluated according to the following procedures 1 to 5.
(Procedure 1) X hit compounds and Y non-hit compounds are mixed with a certain protein (target protein).
(Procedure 2) Invariant AAM descriptors of all (X + Y) compounds are calculated.
(Procedure 3) The similarity of each descriptor is calculated.
(Procedure 4) The (X + Y) compounds are divided into teams based on the similarities of the invariant AAM descriptors.
(Procedure 5) It is checked whether the teams in which hits are collected are
29

mechanically generated.
[0079] As a result of division of 10,933 compounds having 183 hits (a hit content of 1.6%) for the protein ABL1 (kinase) into teams by according to the above-described procedures, the number of teams is 221. A certain team contains 16 hits and 14 other compounds, and the hit content is 53.3%. Further, this team also contains the compound shown in Fig. 12A and the compound shown in Fig. 12B. However, in a case where the fingerprint, which is a descriptor of the related art, is used, the similarity of these compounds is 25%. It is recognized as not being a hit even though it is a hit originally. As described above, in a case where the invariant AAM descriptor according to the first embodiment of the present invention is used in the above-described team, it was found that hits which are not collected in a case of using the fingerprint belong to the same team.
[0080] Fig. 13 shows the results of acquiring the easiness of finding hits (= expectation value; number of hits × hit content) for each of the 221 teams described above. For comparison, the result obtained in a case where the teams are randomly divided and the result obtained in a case where teams are divided using the fingerprint are shown. Based on these results, it was found that in a case where the invariant AAM descriptor is used for the above-described compound group, teams having hits more than the hits of the randomly divided teams or the teams divided using the fingerprint are generated. In Fig. 13, the team numbers vary depending on the team division method (random, the invariant AAM descriptor, the fingerprint), and thus the superiority of the team division is determined not by comparing the expectation values with the same team number but by verifying “whether teams with high expectation values (having more hits) are included or not”. [0081]
Fig. 14 is a graph showing simulation results of Importance Sampling for the above-described compound group. In a case where the invariant AAM descriptor according to the first embodiment is used, the hit search time (the number of times of drug efficacy evaluations for finding the number of same hits) compared with a case of the random team division is reduced to approximately one-half in a case of 50% search and reduced to one-fourth in a case of 25% search. Meanwhile, in a case of the team division using the fingerprint, the hit search time is not shortened. Further, the simulation of the Importance Sampling here indicates a method of providing variables showing the priority for each team and collecting more hits with a smaller number of times of measurements while updating the variables such that the priority of the team from which hits have found is raised and the
30

priority of the team that does not have hits is lowered for each measurement. Fig. 14 shows an example in which an NB method (NB: Naive Bayes), which is a kind of machine learning method is used for controlling the priority, but the present invention is not limited thereto. [0082]
Fig. 15 shows a simulation result of Importance Sampling for 32,464 compounds (having 483 hits) for a protein AA2AR, similar to the case of (part 1). The protein AA2AR is a membrane protein from which an X-ray crystal structure is difficult to obtain, but the hit search time is reduced to approximately one-half in a case of 50% search even with such a membrane protein.
[0083] According to the invariant AAM descriptor described above, since compounds having similar descriptors exhibit similar drug efficacies (binding to a target protein), the chemical properties of the target structure (a compound or a pocket structure) are accurately exhibited. According to the invariant AAM descriptor in which the three-dimensional AAM descriptor is converted into an invariant, the feature quantity is easily handled and the data capacity can be easily reduced while comparison (determination of the drug efficacy) of compounds based on the descriptor is accurately performed by performing conversion into an invariant using the three-dimensional AAM descriptor of two kinds of different amino acids. Further, according to the invariant AAM descriptor, hits are easily found, and the search can be sped up. [0084]
As described above, the screening device 10 according to the first embodiment is capable of calculating the feature quantity (the three-dimensional AAM descriptor or the invariant AAM descriptor) accurately showing the chemical property of the target structure using the feature quantity calculating method and the feature quantity calculating program according to the embodiment of the present invention. [0085]
Extraction of a target compound (pharmaceutical candidate compound) from a plurality of compounds using the above-described three-dimensional AAM descriptor and invariant AAM descriptor will be described. A target compound is extracted in a mode (first mode) in which the extraction is carried out based on the descriptor (the three-dimensional AAM descriptor or the invariant AAM descriptor) of the ligand or in a mode (second mode) in which the extraction is carried out based on the descriptor (the three-dimensional AAM descriptor or the invariant AAM descriptor) of the pocket structure of the target protein. The
31

mode for extraction can be selected from the above-described modes according to the operation of the user via the operation unit 400. [0086]
Fig. 16 is a flowchart showing a procedure for screening using a three-dimensional AAM descriptor of a ligand. After the start of the process, the feature quantity calculation unit 120 calculates a three-dimensional AAM descriptor of the ligand (Step S300: the feature quantity calculating step). Since the ligand is a compound whose binding to the target protein has been confirmed, the three-dimensional AAM descriptor in Step S300 can be calculated according to the procedure shown in the flowchart of Fig. 5.
[0087] As described above with reference to Fig. 4, in the screening device 10, the three-dimensional structure of the compound formed of a plurality of atoms and the three-dimensional AAM descriptor corresponding to the three-dimensional structure are stored in the storage unit 200 in association with each other for each of the plurality of compounds. The similarity calculation unit 130 calculates the similarity between the three-dimensional AAM descriptor of the compound and the three-dimensional AAM descriptor of the ligand calculated in Step S300 (Step S302: the similarity calculating step). After the calculation of the similarity, the compound extraction unit 140 extracts the target compound based on the similarity (Step S304: the target compound extracting step). As described above, in a case where three-dimensional AAM descriptors are similar, since similar drug efficacies (binding to the target protein) are exhibited, a compound having drug efficacy similar to that of the ligand (that is, a target compound serving as a pharmaceutical candidate) can be extracted by using the similarity of the three-dimensional AAM descriptor. Further, the extraction of the target compound based on the similarity (Step S304) can be specifically performed by “extracting a compound having a similarity greater than or equal to the threshold” or “extracting a compound in a descending order of the similarity”.
[0088] Fig. 16 shows the procedure for screening using a three-dimensional AAM descriptor,
but the screening using an invariant AAM descriptor can also be performed in the same
procedure as described above. Specifically, the feature quantity calculation unit 120
calculates the invariant AAM descriptor (the invariant feature quantity) of the ligand according to the procedure of Fig. 5 and Equations (2) and (3), and the similarity calculation unit 130 calculates the similarity between the compound stored in the storage unit 200 and the invariant AAM descriptor. After the calculation of the similarity, the compound extraction unit 140 extracts the target compound based on the similarity. Specifically, the target compound can
32

be extracted based on the similarity in the same manner as the extraction of the three-dimensional AAM descriptor.
[0089] Figs. 17A and 17B are tables showing an example of a screening result of ligand input. Fig. 17A shows the result in a case of “extraction of a compound having a similarity greater than or equal to the threshold” using a three-dimensional AAM descriptor, and Fig. 17B shows the result in a case of "extraction of a compound in a descending order of the similarity" using an invariant AAM descriptor. Further, in Fig. 17A, the compound is extracted based on a three-dimensional AAM descriptor (g1(r)) of the amino acid 1, but the compound may be extracted based on the three-dimensional AAM descriptor (for example, g2(r)) of other amino acids (amino acids 2 to 20). In addition, the similarities (the similarity between the values of g1(r) and the similarity between the values of g2(r)) of a plurality of three-dimensional AAM descriptors (for example, g1(r) and g2(r)) of different amino acids are respectively calculated, and compounds may be extracted based on the results. The number of kinds of the three-dimensional AAM descriptors used for extraction of a compound may be one, but extraction of a compound based on the similarity can be accurately performed using a plurality of kinds of three-dimensional AAM descriptors. Further, in a case where a plurality of kinds of three-dimensional AAM descriptors are used, the combination of amino acids among the descriptors is not particularly limited (for example, a combination of g1(r) and g2(r) or a combination of g3(r) and g4(r) may be used).
[0090] Similarly, in Fig. 17B, a compound is extracted based on an invariant AAM descriptor (F12(s)) of the amino acids 1 and 2, but the combination of amino acids used for calculation of the invariant AAM descriptor may vary (for example, F34(s) with amino acids 3 and 4). Further, a compound may be extracted based on a plurality of invariant AAM descriptors (for example, F12(s) and F34(s)) with different combinations of amino acids (for example, the similarity between the values of F12(s) and the similarity between the values of F34(s) are used). The number of kinds of the invariant AAM descriptors used for extraction of a compound may be one, but extraction of a compound based on the similarity can be accurately performed using a plurality of kinds of invariant AAM descriptors. Further, in a case where a plurality of kinds of invariant AAM descriptors are used, the combination of amino acids among the descriptors is not particularly limited (for example, a combination of F12(s) and F34(s) or a combination of F12(s) and F13(s) may be used). The processing unit 100 (the feature quantity calculation unit 120, the similarity calculation unit 130, and the compound extraction unit 140) may determine which amino acid is to be used for calculation of the descriptor and the
33

similarity according to the user's instruction via the operation unit 400, but the determination may be made by the processing unit 100 regardless of the user's instruction.
[0091] Further, the threshold of the similarity is set to 80% in Fig. 17A, and the number of times of extraction is set to 100 in Fig. 17B, but these values are merely examples. The threshold and the number of times of extraction can be set according to the conditions, for example, the accuracy of screening. The setting can be performed in response to a user input via the operation unit 400. Further, “a compound may be extracted in a descending order of the similarity” in a case where a three-dimensional AAM descriptor is used in contrast to Figs. 17A and 17B, and “a compound having a similarity greater than or equal to the threshold may be extracted” in a case where an invariant AAM descriptor is used. The compound extraction unit 140 stores the extraction result as shown in Figs. 17A and 17B in the storage unit 200 as the compound extraction result 260 (see Fig. 3). [0092]
Fig. 18 is a flowchart showing a procedure for screening of a pocket structure of a target protein using a three-dimensional AAM descriptor. After the start of the process, the feature quantity calculation unit 120 calculates a three-dimensional AAM descriptor of the pocket structure of the target protein (Step S400: the feature quantity calculating step). The three-dimensional AAM descriptor in Step S400 can be calculated according to the procedure shown in the flowchart of Fig. 9. The similarity calculation unit 130 calculates the similarity between the three-dimensional AAM descriptor of the compound and the three-dimensional AAM descriptor of the pocket structure calculated in Step S400 (Step S402: the similarity calculating step). After the calculation of the similarity, the compound extraction unit 140 extracts the target compound based on the similarity (Step S404: the target compound extracting step). Similar to the case of the ligand input described above, the extraction of the target compound based on the similarity (Step S404) can be specifically performed by “extracting a compound having a similarity greater than or equal to the threshold” or “extracting a compound in a descending order of the similarity”.
[0093] Even in a case of using the invariant AAM descriptor, a target compound can be extracted according to the same procedure as in the flowchart of Fig. 18.
[0094] Figs. 19A and 19B are tables showing an example of a screening result of target protein input. Fig. 19A shows the result in a case of “extraction of a compound having a similarity greater than or equal to the threshold” using a three-dimensional AAM descriptor, and Fig. 19B shows the result in a case of "extraction of a compound in a descending order of the
34

similarity" using an invariant AAM descriptor. The threshold of the similarity and the number of times of extraction can be set according to the conditions, for example, the accuracy of screening. The setting can be performed in response to a user input via the operation unit 400. Further, “a compound may be extracted in a descending order of the similarity” in a case where a three-dimensional AAM descriptor is used in contrast to Figs. 19A and 19B, and “a compound having a similarity greater than or equal to the threshold may be extracted” in a case where an invariant AAM descriptor is used.
[0095] In a case of screening for the target protein input, the kind of amino acid may be changed in the same manner as in the case of screening for the ligand input (see Figs. 17A and 17B and the description of these tables), or a plurality of descriptors of different amino acids (the three-dimensional AAM descriptor and the invariant AAM descriptor) may be used. The number of kinds of the descriptors used for extraction of a compound may be one, but extraction of a compound based on the similarity can be accurately performed using a plurality of kinds of descriptors. Further, in a case where a plurality of kinds of descriptors are used, the combination of amino acids among the descriptors is not particularly limited. The processing unit 100 (the feature quantity calculation unit 120, the similarity calculation unit 130, and the compound extraction unit 140) may determine which amino acid is to be used for calculation of the descriptor and the similarity according to the user's instruction via the operation unit 400, but the determination may be made by the processing unit 100 regardless of the user's instruction.
[0096] The compound extraction unit 140 stores the extraction result as shown in Figs. 19A and 19B in the storage unit 200 as the compound extraction result 260 (see Fig. 3). [0097]
As described above, the screening device 10 according to the first embodiment is capable of efficient screening of a pharmaceutical candidate compound according to the screening method and the screening program according to the embodiment of the present invention using the feature quantity (the three-dimensional AAM descriptor or the invariant AAM descriptor) calculated by the feature quantity calculating method and the feature quantity calculating program according to the embodiment of the present invention. [0098]
A compound creating device according to a second embodiment of the present invention will be described. Fig. 20 is a block diagram showing a configuration of a compound creating device 20 (a feature quantity calculating device or a compound creating
35

device). Further, the same elements as those in the first embodiment are denoted by the same reference numerals, and detailed description thereof will not be provided.
[0099] The compound creating device 20 includes a processing unit 101. The processing unit 101 is formed as shown in Fig. 21 and includes an information input unit 110, a feature quantity calculation unit 120 (feature quantity calculation unit), a generator construction unit 132 (generator construction unit), a compound three-dimensional structure generation unit 142 (compound three-dimensional structure generation unit), and a display control unit 150. The functions of the information input unit 110, the feature quantity calculation unit 120, and the display control unit 150 are respectively the same as the information input unit 110, the feature quantity calculation unit 120, and the display control unit 150 in the above-described screening device 10. The functions of these units can be realized using various processors in the same manner as described above in the section of the screening device 10.
[0100] Fig. 22 is a diagram showing information stored in the storage unit 201. The storage unit 201 stores a three-dimensional structure generation result 270 instead of the compound extraction result 260 in the screening device 10. The information stored in the storage unit 201 is stored in association as described above with reference to Fig. 4. [0101]
Generation of a three-dimensional structure of a target compound (pharmaceutical candidate compound) using the above-described three-dimensional AAM descriptor and invariant AAM descriptor will be described. Since search is not performed in the generation of a three-dimensional structure of a target compound using the compound creating device 20, the three-dimensional structure of the compound can be generated even in a case of “no solution was found as the result of screening search”, and thus the three-dimensional structure of the pharmaceutical candidate compound can be efficiently created. A three-dimensional structure is generated in a mode in which the generation is carried out based on the descriptor (the three-dimensional AAM descriptor or the invariant AAM descriptor) of the ligand or in a mode in which the generation is carried out based on the descriptor (the three-dimensional AAM descriptor or the invariant AAM descriptor) of the pocket structure of the target protein. The mode for generation of a three-dimensional structure can be selected from the above-described modes according to the operation of the user via the operation unit 400. [0102]
Fig. 23 is a flowchart showing a procedure for generating a three-dimensional structure in a case of ligand input. After the start of the process, the feature quantity
36

calculation unit 120 calculates a descriptor (a three-dimensional AAM descriptor) of the ligand (Step S500: the target structure designating step, the three-dimensional structure generating step, and the feature quantity calculating step). The process of Step S500 can be performed using the feature quantity calculating method and the feature quantity calculating program according to the embodiment of the present invention in the same manner as in the first embodiment (see Figs. 5 to 8 and the description of these drawings).
[0103] In Step S502, the generator construction unit 132 constructs a generator through machine learning (a generator constructing step). Hereinafter, the process of Step S502 will be described with reference to Fig. 24. (Step 1) The feature quantity calculation unit 120 calculates three-dimensional AAM descriptors of a plurality of compounds and creates a pair (three-dimensional data) of a structural formula (a three-dimensionalized structural formula) and a three-dimensional AAM descriptor. (Step 2) The generator construction unit 132 constructs a generator through machine learning (deep learning) using a three-dimensional structure of a compound as teacher data and a three-dimensional AAM descriptor as an explanatory variable. The method of deep learning is not limited to a specific method, and a simple fully-coupled neural network or a convolutional neural network (CNN) may be employed. However, since the generation accuracy of the three-dimensional structure depends on the learning method to be used, it is preferable to select a learning method according to the condition for generating the three-dimensional structure and the condition such as the required accuracy. As a generator construction method, the methods described in additional notes 14 and 15 described below may be used.
[0104] After the completion of the processes of Steps 1 and 2 described above, the process
returns to the flowchart of Fig. 23. The compound three-dimensional structure generation
unit 142 generates a three-dimensional structure (three-dimensionalized structural formula) of
the target compound (hit) from the three-dimensional AAM descriptor of the ligand using the
constructed generator (Step S504: the compound three-dimensional structure generating step).
In this manner, the three-dimensional structure of a compound having drug efficacy (binding
to a target protein) similar to that of a ligand, that is, a pharmaceutical candidate compound
can be obtained. Further, a plurality of three-dimensional structures that provide the same
three-dimensional AAM descriptor may be present. The compound three-dimensional
structure generation unit 142 stores the generated three-dimensional structure in the storage unit 201 in association with the three-dimensional AAM descriptor (the three-dimensional AAM descriptor 230) as the three-dimensional structure generation result 270 (see Fig. 22).
37

The display control unit 150 may display the generated three-dimensional structure on the monitor 310 in response to the user's instruction via the operation unit 400.
[0105] Further, in the procedure described above, the number of kinds of amino acids used for
calculation of the three-dimensional AAM descriptor for machine learning may be one or
plural. However, the accuracy of the generated three-dimensional structure can be improved
by calculating the three-dimensional AAM descriptor of a plurality of kinds of amino acids
and providing the result for learning. Further, in a case where a plurality of
three-dimensional AAM descriptors having different kinds of amino acids are used, the combination of amino acids among the descriptors is not particularly limited. The processing unit 100 (the feature quantity calculation unit 120, the similarity calculation unit 130, and the compound extraction unit 140) may determine which amino acid is to be used for calculation of the three-dimensional AAM descriptor and for provision for learning according to the user's instruction via the operation unit 400, but the determination may be made by the processing unit 100 regardless of the user's instruction. [0106]
An example of a three-dimensional structure generated using a generator constructed through machine learning will be described. In this example, 1,800 compounds among the library compounds are learned according to the above-described method using a simple fully-coupled neural network, and the extent to which the three-dimensional structures of the remaining 200 compounds can be reproduced is examined. The results are shown in Fig. 25. In a case where the number of interlayers in the neural network is increased, the average cos similarity reaches 59%. Figs. 26A and 26B are diagrams showing a three-dimensional structure (structural formula) and a correct structural formula (Figs. 26A and 26B, respectively) generated from a three-dimensional AAM descriptor as an example showing such similarity.
[0107]
The three-dimensional structure generated according to the above-described
procedure is affected by the features of the compound provided as teacher data. Therefore,
by selecting the features of the compound to be provided as teacher data, a compound having a
three-dimensional structure with different features can be generated. For example, a
compound having drug efficacy similar to that of a ligand and having a three-dimensional structure that is easy to synthesize can be generated by providing, as teacher data, a
38

three-dimensional AAM descriptor of a compound having a three-dimensional structure that is easy to synthesize. It is possible to select which compound to be provided for the three-dimensional AAM descriptor as the teacher data according to the features of the compound intended to be generated. [0108]
In Figs. 23 to 26, the generation of the three-dimensional structure using the three-dimensional AAM descriptor has been described. Meanwhile, similarly to the case of using the three-dimensional AAM descriptor, the three-dimensional structure of the target compound can be generated through machine learning (deep learning) using the invariant AAM descriptor as teacher data and the three-dimensional structure (three-dimensionalized structural formula) as an explanatory variable even in a case of using the invariant AAM descriptor (invariant feature quantity). [0109]
The compound creating device 20 is capable of generating a three-dimensional structure of a target compound by setting a target protein as an input, in addition to the generation of the three-dimensional structure by ligand input. Even in this case, similarly to the case of ligand input, generation of a three-dimensional structure can be performed using a three-dimensional AAM descriptor and generation of a three-dimensional structure can be performed using an invariant AAM descriptor.
[0110] Fig. 27 is a flowchart showing a procedure for generating a three-dimensional structure in a case of setting a target protein as an input (a three-dimensional AAM descriptor is set to be used). In a case where the process is started, the feature quantity calculation unit 120 calculates the three-dimensional AAM descriptor of the pocket structure of the target protein (Step S600: the target structure designating step, the three-dimensional structure generating step, and the feature quantity calculating step). The process of Step S600 can be performed using the feature quantity calculating method according to the embodiment of the present invention, similarly to the first embodiment (see Figs. 9 and 10 and the description of these drawings).
[0111] In Step S602, the generator construction unit 132 constructs a generator through machine learning (deep learning) similar to the case of ligand input (the generator constructing step). The construction of the generator can be performed in the same manner as in Steps 1 and 2 described above. The compound three-dimensional structure generation unit 142 generates a three-dimensional structure (three-dimensionalized structural formula) of the target
39

compound (hit) from the three-dimensional AAM descriptor of the pocket structure using the
constructed generator (Step S604: the compound three-dimensional structure generating step).
In this manner, the three-dimensional structure of a compound having drug efficacy (binding
to a target protein) similar to that of the pocket structure, that is, a pharmaceutical candidate
compound can be obtained. Further, a plurality of three-dimensional structures that provide
the same three-dimensional AAM descriptor may be present. The compound
three-dimensional structure generation unit 142 stores the generated three-dimensional structure in the storage unit 201 in association with the three-dimensional AAM descriptor (the three-dimensional AAM descriptor 230) as the three-dimensional structure generation result 270 (see Fig. 22). The display control unit 150 may display the generated three-dimensional structure on the monitor 310 in response to the user's instruction via the operation unit 400. [0112]
As described above, the compound creating device 20 according to the second embodiment efficiently creates a three-dimensional structure of a pharmaceutical candidate compound according to the feature quantity calculating method and the compound creating program according to the embodiment of the present invention using the feature quantity (the three-dimensional AAM descriptor or the invariant AAM descriptor) calculated using the compound creating method and the feature quantity calculating program according to the embodiment of the present invention. [0113]
The first embodiment described above is an aspect in which the calculation of the
feature quantity and screening based on the calculation are performed, and the second
embodiment is an aspect in which the calculation of the feature quantity and creation of the
three-dimensional structure of the target compound based on the calculation are performed.
In addition to the calculation of the feature quantity, both the screening and the creation of a
three-dimensional structure of the target compound may be performed. Therefore, a
pharmaceutical candidate compound search device 30 (the feature quantity calculating device, the screening device, and the compound creating device; see Fig. 28) according to the third embodiment includes a processing unit 102 shown in Fig. 28 in place of the processing unit 100 of the screening device 10 shown in Fig. 1 or the processing unit 101 of the compound creating device 20 shown in Fig. 20. As shown in Fig. 29, the processing unit 102 includes a feature quantity calculation unit 120 (the feature quantity calculation unit and the invariant conversion unit), a similarity calculation unit 130 (the similarity calculation unit), a generator
40

construction unit 132 (the generator construction unit), a compound extraction unit 140 (the compound extraction unit), and a compound three-dimensional structure generation unit 142 (the compound three-dimensional structure generation unit) and can perform calculation of a feature quantity, screening, and creation of a three-dimensional structure of a compound. In addition, the pharmaceutical candidate compound search device 30 stores information related to the above-described performance in the storage unit 202. Specifically, as shown in Fig. 30, information (see Figs. 3 and 22) stored in the storage unit 200 and the storage unit 201 is stored in the storage unit 202 together.
[0114] Since other elements are the same as those of the screening device 10 shown in Fig. 1 and the compound creating device 20 shown in Fig. 20, the elements are denoted by the same reference numerals and the detailed description thereof will not be provided.
[0115] With the above-described configuration, also in the pharmaceutical candidate compound search device 30 according to the third embodiment, the feature quantity accurately showing the chemical properties of the target structure is calculated, screening of a pharmaceutical candidate compound is efficiently performed, and a three-dimensional structure of the pharmaceutical candidate compound can be efficiently created, similarly to the screening device 10 and the compound creating device 20.
[0116] The embodiments of the present invention have been described above, but the present invention is not limited to the above-described aspects, and various modifications can be made without departing from the spirit of the present invention as exemplified below. [0117]
In the present invention, deoxyribonucleic acid (DNA), ribonucleic acid (RNA), cell
membranes, and polysaccharides can be treated in addition to proteins as the targets of drugs.
However, in the first to third embodiments, it is necessary to change the amino acid to another
one. Specifically, an amino acid is changed to a nucleic acid base in a case of DNA, an
amino acid is changed to a nucleic acid base in a case of RNA, an amino acid is changed to a
lipid molecule in a case of cell membranes, and an amino acid is changed to a monosaccharide
molecule in a case of polysaccharides. In the description below, the reason why DNA, RNA,
cell membranes, and polysaccharides can be treated with this change in the present invention
will be described. Proteins, DNA, RNA, cell membranes, and polysaccharides are
collectively referred to as biopolymers and are made up of unique building blocks. Specifically, the building block of proteins is an amino acid, the building block of DNA is a nucleic acid base, the building block of RNA is similarly a nucleic acid base, the building
41

block of cell membranes is a lipid molecule, and the building block of polysaccharides is a
monosaccharide molecule. Since DNA, RNA, cell membranes, and polysaccharides, which
are biopolymers other than proteins also have pockets that are active sites similarly to the
proteins, even in a case where a drug target (target biopolymer) is DNA, RNA, cell
membranes and polysaccharides, the present invention can deal with this by changing amino
acids to the building blocks of the targets in the first to third embodiments shown in the case of
proteins. Further, water can be considered in a case of quantifying the degree of
accumulation of amino acids, nucleic acid bases, lipid molecules, and monosaccharide molecules in the periphery of a compound or a pocket structure. [0118]
In the present invention, in addition to the typical activity which is the "activity of a target biomolecule alone by a compound", the "activity of a cell, which is a composite formed of other biomolecules in addition to the target biomolecule by a compound" can also be treated.
[0119] <(Modification Example 1) feature quantity of biopolymer other than amino acid and use thereof>

In a case of treating DNA, RNA, cell membranes, and polysaccharides, which are biopolymers (compounds) other than proteins as drug targets (target biopolymers), the probe used for calculation of the feature quantity is not an amino acid but a different substance (the building block of each target). Specifically, in a case where the targets are “DNA, RNA, cell membranes, and polysaccharides”, the probes are respectively set as “one or more kinds of nucleic acid bases, one or more kinds of nucleic acid bases, one or more kinds of lipid molecules, and one or more kinds of monosaccharide molecules”. Further, water and one or more kinds of ions may be considered during the quantification of the degree of accumulation using these as probes. Further, in a case where the targets are formed of a plurality of kinds of biopolymers among "DNA, RNA, cell membranes, and polysaccharides", the probe can also be set as one or more of "one or more kinds of nucleic acid bases, one or more kinds of nucleic acid bases, one or more kinds of lipid molecules, one or more kinds of monosaccharide molecules, water, and one or more kinds of ions” (the kind, the number, and the combination thereof may be optional according to the configurations of the targets) in accordance of the configurations of the targets. The ions constituting the probes may be monatomic ions or polyatomic ions. Further, all the probes are assumed to generate van der Waals forces.
42

[0120]
The configurations of devices for calculating the feature quantity (the feature quantity according to Modification Example 1) and performing screening (the feature quantity calculating device and the screening device) are the same as those in the first embodiment (see Figs. 1 to 3). However, the descriptor according to Modification Example 1 is calculated and stored instead of the three-dimensional AAM descriptor 230 in Fig. 3, and the invariant feature quantity according to Modification Example 1 is calculated and stored instead of the invariant AAM descriptor 240. The procedures for calculation of the feature quantity and screening are the same as those in the first embodiment, and the feature quantity calculating method, the feature quantity calculating program, the screening method, and the screening program according to the embodiment of the present invention can be used. Specifically, the invariant feature quantity according to Modification Example 1 is calculated (see Equation (2)) by calculating the distribution function (see Equation (1)) using “one or more (the kind, the number, and the combination thereof may be optional) selected from one or more kinds of nucleic acid bases, one or more kinds of lipid molecules, one or more kinds of monosaccharide molecules, water, and one or more kinds of ions” instead of the “amino acid” as the probe in the calculation of the feature quantity (see Fig. 5) according to the first embodiment, calculating the feature quantity according to Modification Example 1 from the distribution function, and performing Fourier transform on the feature quantity according to Modification Example 1. Further, the invariant feature quantity according to Modification Example 1 may be calculated by angular integration of the correlation function (see Equations (3) and (4)) using the feature quantity of two different kinds of probes according to Modification Example 1 (the first probe formed of one or more selected from one or more kinds of nucleic acid bases, one or more kinds of lipid molecules, one or more kinds of monosaccharide molecules, water, and one or more kinds of ions and the second probe that is formed of one or more selected from one or more kinds of nucleic acid bases, one or more kinds of lipid molecules, one or more kinds of monosaccharide molecules, water, and one or more kinds of ions and different from the first probe).
[0121] In addition, the target compound can be extracted based on the similarity between the feature quantity of according to Modification Example 1 a plurality of compounds and the feature quantity according to Modification Example 1 of the binding compound using the feature quantity according to Modification Example 1 instead of the three-dimensional AAM descriptor in the first embodiment. A compound having a similarity greater than or equal to
43

the threshold may be extracted, or a compound may be extracted in a descending order of the
similarity.
[0122]
The configurations of devices for calculating a feature quantity (the feature quantity according to Modification Example 1) and creating a compound (the feature quantity calculating device and the compound creating device) are the same as those in the second embodiment (see Figs. 20 to 22). However, the descriptor according to Modification Example 1 (the feature quantity according to Modification Example 1) is calculated and stored instead of the three-dimensional AAM descriptor 230 in Fig. 22, and the invariant feature quantity according to Modification Example 1 is calculated and stored instead of the invariant AAM descriptor 240. The procedures for calculation of the feature quantity and creation of a compound are the same as those in the second embodiment, and the feature quantity calculating method, the feature quantity calculating program, the compound creating method, and the compound creating program according to the embodiment of the present invention can be used. In Modification Example 1, a generator is constructed through machine learning (deep learning) using the three-dimensional structure of the compound as teacher data and the feature quantity according to Modification Example 1 as an explanatory variable, and the three-dimensional structure of the target biopolymer can be generated from the feature quantity according to Modification Example 1 of the binding compound that is a compound whose binding to the target biopolymer has been confirmed using the constructed generator. Further, similar to the first to third embodiments, a compound having a three-dimensional structure with different features can be generated by selecting features of a compound to be provided as teacher data.
[0123] In the same manner as in the first to third embodiments described above, since the drug
efficacy of a compound (the binding force with respect to the target such as DNA) is locally
exhibited as the result of an interaction between a compound and a nucleic acid base (probe),
in a case where the degree of accumulation of nucleic acid bases and the like is similar
between compounds, the compounds have similar binding forces with respect to the targets.
That is, the compounds with similar feature quantities according to Modification Example 1
exhibit similar drug efficacies. Therefore, in Modification Example 1, the chemical
properties of the compound can be accurately exhibited by the feature quantity according to Modification Example 1. In addition, in a case where the feature quantities according to Modification Example 1 are similar between the target compound and the binding compound
44

that is bound to the target biopolymer, the drug efficacies of both the binding compound and the target compound are similar. Therefore, according to Modification Example 1, a target compound having drug efficacy similar to that of the binding compound is extracted based on the feature quantity according to Modification Example 1 so that screening of a pharmaceutical candidate compound can be efficiently performed. Further, according to Modification Example 1, similarly to the above-described embodiments, a structural formula of a compound having a feature quantity similar to the feature quantity (the feature quantity according to Modification Example 1) of the binding compound (accordingly, the drug efficacies are similar) is generated without performing search, and thus the three-dimensional structure of the pharmaceutical candidate compound can be efficiently created.
[0124] In addition, in a case where the calculation of a feature quantity, the screening, and the creation of a compound are performed (corresponding to the third embodiment), the calculation, the screening, and the creation can be performed using the same configurations as in Figs. 28 to 30. [0125] <(Modification Example 2) feature quantity of amino acid and others and use thereof>

In Modification Example 2, “composite of proteins and biopolymers other than the proteins (DNA, RNA, cell membranes, and polysaccharides)” are set as targets. Further, “one or more kinds of amino acids” (first probe) and “one or more selected from one or more kinds of nucleic acid bases, one or more kinds of nucleic acid bases, one or more kinds of lipid molecules, one or more kinds of monosaccharide molecules, water, and one or more kinds of ions" (second probe; the kind, the number, and the combination thereof may be optional) are set as probes. The configurations of the first and second probes (the kind, the number, and the combination thereof) can be set according to the configurations of the targets. The ions constituting the probes may be monatomic ions or polyatomic ions. Further, all the probes are assumed to generate van der Waals forces. [0126]
The configurations of devices for calculating the feature quantity (the feature quantity according to Modification Example 2) and performing screening (the feature quantity calculating device and the screening device) are the same as those in the first embodiment (see Figs. 1 to 3). However, a descriptor according to Modification Example 2 (the feature quantity according to Modification Example 2) is calculated and stored instead of the three-dimensional AAM descriptor 230 in Fig. 3, and an invariant feature quantity according
45

to Modification Example 2 is calculated and stored instead of the invariant AAM descriptor 240. The procedures for calculation of the feature quantity and screening are the same as those in the first embodiment, and the feature quantity calculating method, the feature quantity calculating program, the screening method, and the screening program according to the embodiment of the present invention can be used. Specifically, the invariant feature quantity according to Modification Example 2 is calculated (see Equation (2)) by calculating the distribution function (see Equation (1)) using “one or more kinds of amino acids” (the first probe) and “one or more selected from one or more kinds of nucleic acid bases, one or more kinds of lipid molecules, water, one or more kinds of monosaccharide molecules, and one or more kinds of ions” (the second probe; the kind, the number, and the combination thereof may be optional) instead of the “amino acid” as the probe in the calculation of the feature quantity according to Modification Example 2 (see Fig. 5), calculating the feature quantity according to Modification Example 2 from the distribution function, and performing Fourier transform on the feature quantity according to Modification Example 2. Further, the invariant feature quantity according to Modification Example 2 may be calculated by angular integration of the correlation function (see Equations (3) and (4)) using the feature quantity according to Modification Example 2 of two kinds of probes in which at least one of “one or more kinds of amino acids” (first probe) or “one or more selected from one or more kinds of nucleic acid bases, one or more kinds of lipid molecules, one or more kinds of monosaccharide molecules, water, and one or more kinds of ions” (second probe; the kind, the number, and the combination thereof may be optional) is different.
[0127] In addition, the target compound can be extracted based on the similarity between the feature quantity according to Modification Example 2 of a plurality of compounds and the feature quantity according to Modification Example 2 of the binding compound using the feature quantity according to Modification Example 2 instead of the three-dimensional AAM descriptor in the first embodiment. A compound having a similarity greater than or equal to the threshold may be extracted, or a compound may be extracted in a descending order of the similarity. [0128]
The configurations of devices for calculating a feature quantity (the feature quantity according to Modification Example 2) and creating a compound (the feature quantity calculating device and the compound creating device) are the same as those in the second embodiment (see Figs. 20 to 22). However, the descriptor according to Modification
46

Example 2 (the feature quantity according to Modification Example 2) is calculated and stored instead of the three-dimensional AAM descriptor 230 in Fig. 22, and the invariant feature quantity according to Modification Example 2 is calculated and stored instead of the invariant AAM descriptor 240. The procedures for calculation of the feature quantity and creation of a compound are the same as those in the second embodiment, and the feature quantity calculating method, the feature quantity calculating program, the compound creating method, and the compound creating program according to the embodiment of the present invention can be used. In Modification Example 2, a generator is constructed through machine learning (deep learning) using the three-dimensional structure of the compound as teacher data and the feature quantity according to Modification Example 2 as an explanatory variable, and the three-dimensional structure of the target biopolymer can be generated from the feature quantity according to Modification Example 2 of the binding compound that is a compound whose binding to the target biopolymer has been confirmed using the constructed generator. Further, similar to the first to third embodiments and Modification Example 1, a compound having a three-dimensional structure with different features can be generated by selecting features of a compound to be provided as teacher data.
[0129] In the same manner as in the first to third embodiments and Modification Example 1 described above, since the drug efficacy of a compound (the binding force with respect to the target) is locally exhibited as the result of an interaction between a compound and a probe, in a case where the degree of accumulation of probes is similar between compounds, the compounds have similar binding forces with respect to the targets. That is, the compounds with similar feature quantities according to Modification Example 2 exhibit similar drug efficacies. Therefore, in Modification Example 2, the chemical properties of the compound can be accurately exhibited by the feature quantity according to Modification Example 2. In addition, in a case where the feature quantities according to Modification Example 2 are similar between the target compound and the binding compound that is bound to the target biopolymer, the drug efficacies of both the binding compound and the target compound are similar. Therefore, according to Modification Example 2, a target compound having drug efficacy similar to that of the binding compound is extracted based on the feature quantity according to Modification Example 2 so that screening of a pharmaceutical candidate compound can be efficiently performed. Further, according to Modification Example 2, similarly to the above-described embodiments and Modification Example 1, a structural formula of a compound having a feature quantity similar to the feature quantity (the feature
47

quantity according to Modification Example 2) of the binding compound (accordingly, the
drug efficacies are similar) is generated without performing search, and thus the
three-dimensional structure of the pharmaceutical candidate compound can be efficiently
created.
[0130] In addition, in a case where the calculation of a feature quantity, the screening, and the
creation of a compound are performed (corresponding to the third embodiment), the
calculation, the screening, and the creation can be performed using the same configurations as
in Figs. 28 to 30.
[0131] <(Modification Example 3) feature quantity of virtual point electric charge and the like
and use thereof>

In Modification Example 3, a biopolymer (compound) is used as the target, and “one or more selected from a first point electric charge having an electric charge of +1, a second point electric charge having an electric charge of -1, a third point electric charge having an electric charge of +0.1, a fourth point electric charge having an electric charge of -0.1, and a fifth point electric charge having an electric charge of 0 (the kind, the number, and the combination thereof may be optional)” (virtual point electric charges having a real electric charge and generating a van der Waals force) are used as the probes. These point electric charges are an example of the “single point”. [0132]
The configurations of devices for calculating the feature quantity (the feature quantity according to Modification Example 3; the second feature quantity and the second invariant feature quantity) and performing screening (the feature quantity calculating device and the screening device) are the same as those in the first embodiment (see Figs. 1 to 3). However, a second descriptor (the second feature quantity) is calculated and stored instead of the three-dimensional AAM descriptor 230 in Fig. 3, and a second invariant feature quantity is calculated and stored instead of the invariant AAM descriptor 240. The procedures for calculation of the feature quantity and screening are the same as those in the first embodiment, and the feature quantity calculating method, the feature quantity calculating program, the screening method, and the screening program according to the embodiment of the present invention can be used. Specifically, the second invariant feature quantity is calculated (see Equation (2)) by calculating the distribution function (see Equation (1)) using “one or more selected from a first point electric charge having an electric charge of +1, a second point
48

electric charge having an electric charge of -1, a third point electric charge having an electric charge of +0.1, a fourth point electric charge having an electric charge of -0.1, and a fifth point electric charge having an electric charge of 0 (the kind, the number, and the combination thereof may be optional)” instead of the “amino acid” as the probe in the calculation of the feature quantity according to the first embodiment (see Fig. 5), calculating the second feature quantity from the distribution function, and performing Fourier transform on the second feature quantity. Further, the second invariant feature quantity may be calculated by angular integration of the correlation function (see Equations (3) and (4)) using the second feature quantity of two different kinds of probes (the first probe formed of one or more selected from the first point electric charge, the second point electric charge, the third point electric charge, the fourth point electric charge, and the fifth point electric charge and the second probe that is formed of one or more selected from the first point electric charge, the second point electric charge, the third point electric charge, the fourth point electric charge, and the fifth point electric charge and different from the first probe).
[0133] In addition, the target compound can be extracted based on the similarity between the second feature quantity of a plurality of compounds and the second feature quantity of the binding compound using the second feature quantity instead of the three-dimensional AAM descriptor in the first embodiment. A compound having a similarity greater than or equal to the threshold may be extracted, or a compound may be extracted in a descending order of the similarity. [0134]
The configurations of devices for calculating a feature quantity (the second feature quantity) and creating a compound (the feature quantity calculating device and the compound creating device) are the same as those in the second embodiment (see Figs. 20 to 22). However, the second descriptor (the second feature quantity) is calculated and stored instead of the three-dimensional AAM descriptor 230 in Fig. 22, and the second invariant feature quantity is calculated and stored instead of the invariant AAM descriptor 240. The procedures for calculation of the feature quantity and creation of a compound are the same as those in the second embodiment, and the feature quantity calculating method, the feature quantity calculating program, the compound creating method, and the compound creating program according to the embodiment of the present invention can be used. In Modification Example 3, a generator is constructed through machine learning (deep learning) using the three-dimensional structure of the compound as teacher data and the second feature quantity as
49

an explanatory variable, and the three-dimensional structure of the target biopolymer can be generated from the second feature quantity of the binding compound that is a compound whose binding to the target biopolymer has been confirmed using the constructed generator. Further, similar to the first to third embodiments and Modification Examples 1 and 2, a compound having a three-dimensional structure with different features can be generated by selecting features of a compound to be provided as teacher data.
[0135] In the same manner as in the first to third embodiments and Modification Examples 1 and 2 described above, since the drug efficacy of a compound (the binding force with respect to the target) is locally exhibited as the result of an interaction between a compound and a probe, in a case where the degree of accumulation of probes is similar between compounds, the compounds have similar binding forces with respect to the targets. That is, compounds having similar second feature quantities exhibit similar drug efficacies. Therefore, in Modification Example 3, the chemical properties of the compound can be accurately exhibited by the second feature quantity. In addition, in a case where the second feature quantities are similar between the target compound and the binding compound that is bound to the target biopolymer, the drug efficacies of both the binding compound and the target compound are similar. Therefore, according to Modification Example 3, a target compound having drug efficacy similar to that of the binding compound is extracted based on the second feature quantity so that screening of a pharmaceutical candidate compound can be efficiently performed. Further, according to Modification Example 3, similarly to the above-described first to third embodiments and Modification Examples 1 and 2, a structural formula of a compound having a feature quantity similar to the feature quantity (the second feature quantity) of the binding compound (accordingly, the drug efficacies are similar) is generated without performing search, and thus the three-dimensional structure of the pharmaceutical candidate compound can be efficiently created.
[0136] In addition, in a case where the calculation of a feature quantity, the screening, and the creation of a compound are performed (corresponding to the third embodiment), the calculation, the screening, and the creation can be performed using the same configurations as in Figs. 28 to 30.
[0137] <(Modification Example 4) feature quantity of amino acid and virtual point electric charge and use thereof>

In Modification Example 4, a biopolymer (compound) is used as the target, and “a
50

first probe that is one or more kinds of amino acids and a second probe that is one or more selected from a first point electric charge having an electric charge of +1, a second point electric charge having an electric charge of -1, a third point electric charge having an electric charge of +0.1, a fourth point electric charge having an electric charge of -0.1, a dipole in which the first point electric charge and the second point electric charge are disposed to be separated from each other, and a fifth point electric charge having an electric charge of 0 (the kind, the number, and the combination thereof may be optional)” are used as the probes. The second probe may be formed of point electric charges (one or more of the first to fifth point electric charges) excluding the dipole. The first to fifth point electric charges are an example of the “single point”. [0138]
The configurations of devices for calculating the feature quantity (the feature quantity according to Modification Example 4) and performing screening (the feature quantity calculating device and the screening device) are the same as those in the first embodiment (see Figs. 1 to 3). However, a descriptor according to Modification Example 4 (the feature quantity according to Modification Example 4) is calculated and stored instead of the three-dimensional AAM descriptor 230 in Fig. 3, and an invariant feature quantity according to Modification Example 4 is calculated and stored instead of the invariant AAM descriptor 240. The procedures for calculation of the feature quantity and screening are the same as those in the first embodiment, and the feature quantity calculating method, the feature quantity calculating program, the screening method, and the screening program according to the embodiment of the present invention can be used. Specifically, the invariant feature quantity according to Modification Example 4 is calculated (see Equation (2)) by calculating the distribution function (see Equation (1)) using “a first probe that is formed of one or more kinds of amino acids and a second probe that is formed of one or more selected from a first point electric charge having an electric charge of +1, a second point electric charge having an electric charge of -1, a third point electric charge having an electric charge of +0.1, a fourth point electric charge having an electric charge of -0.1, a dipole in which the first point electric charge and the second point electric charge are disposed to be separated from each other, and a fifth point electric charge having an electric charge of 0 (the kind, the number, and the combination thereof may be optional)” instead of the “amino acid” as the probe in the calculation of the feature quantity according to the first embodiment (see Fig. 5), calculating the feature quantity according Modification Example 4 from the distribution function, and
51

performing Fourier transform on the feature quantity according Modification Example 4. The second probe may be formed of point electric charges (one or more of the first to fifth point electric charges) excluding the dipole. Further, the invariant feature quantity according Modification Example 4 may be calculated by angular integration of the correlation function (see Equations (3) and (4)) using the feature quantity according Modification Example 4 of two kinds of probes in which at least one of the first probe or the second probe is different. [0139] In addition, the target compound can be extracted based on the similarity between the feature quantity according Modification Example 4 of a plurality of compounds and the feature quantity according Modification Example 4 of the binding compound using the feature quantity according Modification Example 4 instead of the three-dimensional AAM descriptor in the first embodiment. A compound having a similarity greater than or equal to the threshold may be extracted, or a compound may be extracted in a descending order of the similarity. [0140]
The configurations of devices for calculating a feature quantity (the feature quantity according Modification Example 4) and creating a compound (the feature quantity calculating device and the compound creating device) are the same as those in the second embodiment (see Figs. 20 to 22). However, the descriptor according to Modification Example 4 (the feature quantity according to Modification Example 4) is calculated and stored instead of the three-dimensional AAM descriptor 230 in Fig. 22, and the invariant feature quantity according to Modification Example 4 is calculated and stored instead of the invariant AAM descriptor 240. The procedures for calculation of the feature quantity and creation of a compound are the same as those in the second embodiment, and the feature quantity calculating method, the feature quantity calculating program, the compound creating method, and the compound creating program according to the embodiment of the present invention can be used. In Modification Example 4, a generator is constructed through machine learning (deep learning) using the three-dimensional structure of the compound as teacher data and the feature quantity according to Modification Example 4 as an explanatory variable, and the three-dimensional structure of the target biopolymer can be generated from the feature quantity according to Modification Example 4 of the binding compound that is a compound whose binding to the target biopolymer has been confirmed using the constructed generator. Further, similar to the first to third embodiments, a compound having a three-dimensional structure with different features can be generated by selecting features of a compound to be provided as teacher data.
52

[0141] In the same manner as in the first to third embodiments, since the drug efficacy of a
compound (the binding force with respect to the target) is locally exhibited as the result of an
interaction between a compound and a probe, in a case where the degree of accumulation of
probes is similar between compounds, the compounds have similar binding forces with respect
to the targets. That is, the compounds with similar feature quantities according to
Modification Example 4 exhibit similar drug efficacies. Therefore, in Modification Example 4, the chemical properties of the compound can be accurately exhibited by the feature quantity according to Modification Example 4. In addition, in a case where the feature quantities according to Modification Example 4 are similar between the target compound and the binding compound that is bound to the target biopolymer, the drug efficacies of both the binding compound and the target compound are similar. Therefore, according to Modification Example 4, a target compound having drug efficacy similar to that of the binding compound is extracted based on the feature quantity according to Modification Example 4 so that screening of a pharmaceutical candidate compound can be efficiently performed. Further, according to Modification Example 4, similarly to the above-described embodiments, a structural formula of a compound having a feature quantity similar to the feature quantity (the feature quantity according to Modification Example 4) of the binding compound (accordingly, the drug efficacies are similar) is generated without performing search, and thus the three-dimensional structure of the pharmaceutical candidate compound can be efficiently created.
[0142] In addition, in a case where the calculation of a feature quantity, the screening, and the creation of a compound are performed (corresponding to the third embodiment), the calculation, the screening, and the creation can be performed using the same configurations as in Figs. 28 to 30.
[0143] <(Modification Example 5) feature quantity of nucleic acid base and virtual point electric charge and use thereof>

In Modification Example 5, a biopolymer (compound) is used as the target, and a first probe “that is one or more selected from one or more kinds of nucleic acid bases, one or more kinds of lipid molecules, water, one or more kinds of monosaccharide molecules, and one or more kinds of ions (the kind, the number, and the combination thereof may be optional)” and a second probe “that is one or more selected from a first point electric charge having an electric charge of +1, a second point electric charge having an electric charge of -1, a third point electric charge having an electric charge of +0.1, a fourth point electric charge having an
53

electric charge of -0.1, a dipole in which the first point electric charge and the second point electric charge are disposed to be separated from each other, and a fifth point electric charge having an electric charge of 0 (the kind, the number, and the combination thereof may be optional)” are used as the probes. The first probe may be formed of one or more kinds of monatomic ions, and the second probe may be formed of point electric charges (one or more of the first to fifth point electric charges). The “monatomic ions” and the “point electric charges” are an example of the “single point”. [0144]
The configurations of devices for calculating the feature quantity (the feature quantity according to Modification Example 5; the third feature quantity and the third invariant feature quantity) and performing screening (the feature quantity calculating device and the screening device) are the same as those in the first embodiment (see Figs. 1 to 3). However, a third descriptor (third feature quantity) is calculated and stored instead of the three-dimensional AAM descriptor 230 in Fig. 3, and a third invariant feature quantity is calculated and stored instead of the invariant AAM descriptor 240. The procedures for calculation of the feature quantity and screening are the same as those in the first embodiment, and the feature quantity calculating method, the feature quantity calculating program, the screening method, and the screening program according to the embodiment of the present invention can be used. Specifically, the third invariant feature quantity is calculated (see Equation (2)) by calculating the distribution function (see Equation (1)) using the first probe “that is formed of one or more selected from one or more kinds of nucleic acid bases, one or more kinds of lipid molecules, water, one or more kinds of monosaccharide molecules, and one or more kinds of ions (the kind, the number, and the combination thereof may be optional)” and the second probe “that is formed of one or more selected from a first point electric charge having an electric charge of +1, a second point electric charge having an electric charge of -1, a third point electric charge having an electric charge of +0.1, a fourth point electric charge having an electric charge of -0.1, a dipole in which the first point electric charge and the second point electric charge are disposed to be separated from each other, and a fifth point electric charge having an electric charge of 0 (the kind, the number, and the combination thereof may be optional)”instead of the “amino acid” as the probe in the calculation of the feature quantity according to the first embodiment (see Fig. 5), calculating the third feature quantity from the distribution function, and performing Fourier transform on the third feature quantity. Further, the third invariant feature quantity may be calculated by angular integration of the correlation function (see
54

Equations (3) and (4)) using the third feature quantity of two kinds of probes in which at least one of the first probe or the second probe is different. As described above, the first probe may be formed of one or more kinds of monatomic ions, and the second probe may be formed of point electric charges (one or more of the first to fifth point electric charges). The “monatomic ions” and the “point electric charges” are an example of the “single point”. [0145] In addition, the target compound can be extracted based on the similarity between the third feature quantity of a plurality of compounds and the third feature quantity of the binding compound using the third feature quantity instead of the three-dimensional AAM descriptor in the first embodiment. A compound having a similarity greater than or equal to the threshold may be extracted, or a compound may be extracted in a descending order of the similarity. [0146]
The configurations of devices for calculating a feature quantity (third feature quantity) and creating a compound (the feature quantity calculating device and the compound creating device) are the same as those in the second embodiment (see Figs. 20 to 22). However, the third descriptor (third feature quantity) is calculated and stored instead of the three-dimensional AAM descriptor 230 in Fig. 22, and the third invariant feature quantity is calculated and stored instead of the invariant AAM descriptor 240. The procedures for calculation of the feature quantity and creation of a compound are the same as those in the second embodiment, and the feature quantity calculating method, the feature quantity calculating program, the compound creating method, and the compound creating program according to the embodiment of the present invention can be used. In Modification Example 5, a generator is constructed through machine learning (deep learning) using the three-dimensional structure of the compound as teacher data and the third feature quantity as an explanatory variable, and the three-dimensional structure of the target biopolymer can be generated from the third feature quantity of the binding compound that is a compound whose binding to the target biopolymer has been confirmed using the constructed generator. Further, similar to the first to third embodiments, a compound having a three-dimensional structure with different features can be generated by selecting features of a compound to be provided as teacher data.
[0147] In the same manner as in the first to third embodiments, since the drug efficacy of a compound (the binding force with respect to the target) is locally exhibited as the result of an interaction between a compound and a probe, in a case where the degree of accumulation of probes is similar between compounds, the compounds have similar binding forces with respect
55

to the targets. That is, compounds having similar third feature quantities exhibit similar drug efficacies. Therefore, in Modification Example 5, the chemical properties of the compound can be accurately exhibited by the third feature quantity. In addition, in a case where the third feature quantities are similar between the target compound and the binding compound that is bound to the target biopolymer, the drug efficacies of both the binding compound and the target compound are similar. Therefore, according to Modification Example 5, a target compound having drug efficacy similar to that of the binding compound is extracted based on the third feature quantity so that screening of a pharmaceutical candidate compound can be efficiently performed. Further, according to Modification Example 5, similarly to the above-described embodiments, a structural formula of a compound having a feature quantity similar to the feature quantity (the third feature quantity) of the binding compound (accordingly, the drug efficacies are similar) is generated without performing search, and thus the three-dimensional structure of the pharmaceutical candidate compound can be efficiently created.
[0148] In addition, in a case where the calculation of a feature quantity, the screening, and the creation of a compound are performed (corresponding to the third embodiment), the calculation, the screening, and the creation can be performed using the same configurations as in Figs. 28 to 30.

We Claim:

1.A feature quantity calculating method comprising:
a target structure designating step of designating a target structure formed of a plurality of unit structures having chemical properties; and
a feature quantity calculating step of calculating a feature quantity obtained by quantifying, in a three-dimensional space, a degree of accumulation of one or more kinds of probes in a periphery of a three-dimensional structure of the target structure and calculating the feature quantity from the target structure using a generator formed through machine learning,
wherein the probe is a structure in which a plurality of points having a real electric charge and generating a van der Waals force are disposed to be separated from each other.
2. The feature quantity calculating method according to claim 1,
wherein a compound is designated as the target structure in the target structure designating step, and
a first feature quantity which is a feature quantity obtained by quantifying, in the three-dimensional space, a degree of accumulation of amino acids as the probes in the periphery of the three-dimensional structure of the compound is calculated using the generator in the feature quantity calculating step.
3. The feature quantity calculating method according to claim 1,
wherein a pocket structure bound to a pocket that is an active site of a target protein is designated as the target structure in the target structure designating step, and
a second feature quantity which is a feature quantity obtained by quantifying, in the three-dimensional space, a degree of accumulation of amino acids as the probes in the periphery of a three-dimensional structure of the pocket structure formed of a plurality of virtual spheres is calculated using the generator in the feature quantity calculating step.
4. The feature quantity calculating method according to claim 1,
wherein a compound is designated as the target structure in the target structure designating step, and
a third feature quantity which is a feature quantity obtained by quantifying, in the
113

three-dimensional space, a degree of accumulation of the probes in the periphery of the three-dimensional structure of the compound which is the degree of accumulation using one or more selected from one or more kinds of nucleic acid bases, one or more kinds of lipid molecules, one or more kinds of monosaccharide molecules, water, or one or more kinds of ions formed of a plurality of atoms, as the probes is calculated using the generator in the feature quantity calculating step.
5. A non-transitory computer-readable recording medium which causes a computer to
execute
the feature quantity calculating program according to any one of claims 1 to 4 in a case where a command stored in the recording medium is read by the computer.
6. A feature quantity calculating device comprising:
a target structure designation unit which designates a target structure formed of a plurality of unit structures having chemical properties; and
a feature quantity calculation unit which calculates a feature quantity obtained by quantifying, in a three-dimensional space, a degree of accumulation of one or more kinds of probes in a periphery of a three-dimensional structure of the target structure and calculating the feature quantity from the target structure using a generator formed through machine learning,
wherein the probe is a structure in which a plurality of points having a real electric charge and generating a van der Waals force are disposed to be separated from each other.
7. A screening method of extracting a target compound which is bound to a target protein
from a plurality of compounds, the method comprising:
a storing step of storing a three-dimensional structure of a compound formed of a plurality of atoms and the first feature quantity of the three-dimensional structure of the compound which is calculated using the feature quantity calculating method according to claim 2 in association with each other for each of the plurality of compounds;
a feature quantity calculating step of calculating the first feature quantity of a ligand that is a compound whose binding to the target protein has been confirmed using the feature quantity calculating method according to claim 2;
a similarity calculating step of calculating a similarity between the first feature
114

quantity of the plurality of compounds and the first feature quantity of the ligand; and
a compound extracting step of extracting the target compound from the plurality of compounds based on the similarity.
8. A screening method of extracting a target compound which is bound to a target protein
from a plurality of compounds, the method comprising:
a storing step of storing a three-dimensional structure of a compound formed of a plurality of atoms and the first feature quantity calculated using the feature quantity calculating method according to claim 2 in association with each other for each of the plurality of compounds;
a feature quantity calculating step of calculating the second feature quantity of the pocket structure of the target protein using the feature quantity calculating method according to claim 3;
a similarity calculating step of calculating a similarity between the first feature quantity of the plurality of compounds and the second feature quantity of the pocket structure; and
a compound extracting step of extracting the target compound from the plurality of compounds based on the similarity.
9. A screening method of extracting a target compound which is bound to a target
biopolymer other than a protein from a plurality of compounds, the method comprising:
a storing step of storing a three-dimensional structure of a compound formed of a plurality of atoms and the third feature quantity of the three-dimensional structure of the compound which is calculated using the feature quantity calculating method according to claim 4 in association with each other for each of the plurality of compounds;
a feature quantity calculating step of calculating the third feature quantity of a binding compound that is a compound whose binding to the target biopolymer other than the protein has been confirmed using the feature quantity calculating method according to claim 4;
a similarity calculating step of calculating a similarity between the third feature quantity of the plurality of compounds and the third feature quantity of the binding compound; and
a compound extracting step of extracting the target compound from the plurality of compounds based on the similarity.
115

10. A screening method comprising:
an input step of inputting information related to three-dimensional structures of a plurality of compounds;
a feature quantity calculating step of calculating the feature quantity of each of the plurality of compounds using the feature quantity calculating method according to any one of claims 1 to 4 based on the information;
a characteristic value calculating step of calculating a physical property value indicating a physical property and/or a toxicity value indicating toxicity of each of the plurality of compounds based on the information; and
an extracting step of extracting compounds in which a similarity between the feature quantities satisfies a target value and the physical property value and/or the toxicity value satisfies a target value, from the plurality of compounds.
11. A screening method comprising:
an input step of inputting information related to three-dimensional structures of a plurality of compounds;
a first feature quantity calculating step of calculating the feature quantity of each of the plurality of compounds using the feature quantity calculating method according to any one of claims 1 to 4 based on the information;
a second feature quantity calculating step of calculating a structural feature quantity indicating a compound structure of each of the plurality of compounds based on the information; and
an extracting step of extracting compounds in which a similarity between the feature quantities satisfies a target value and a similarity between the structural feature quantities is less than or equal to a threshold, from the plurality of compounds.
12. A screening method comprising:
an input step of inputting information related to three-dimensional structures of a plurality of compounds;
a feature quantity calculating step of calculating the feature quantity of each of the plurality of compounds using the feature quantity calculating method according to any one of claims 1 to 4 based on the information;
a predicting step of predicting a binding force using a prediction model that outputs
116

the binding force between each of the plurality of compounds and another compound based on the feature quantity; and
an extracting step of extracting compounds in which the binding force satisfies a target value, from the plurality of compounds.
13. A screening method comprising:
an input step of inputting information related to three-dimensional structures of a plurality of compounds;
a setting step of setting a target value of a binding force between each of the plurality of compounds and another compound;
an estimating step of estimating the feature quantity corresponding to the target value using a prediction model that outputs the binding force based on the feature quantity calculated by the feature quantity calculating method according to any one of claims 1 to 4;
a feature quantity calculating step of calculating the feature quantity of each of the plurality of compounds using the feature quantity calculating method according to any one of claims 1 to 4 based on the information; and
an extracting step of extracting compounds in which a similarity between the calculated feature quantity and the estimated feature quantity is greater than or equal to a threshold, from the plurality of compounds based on the calculation results.
14. A non-transitory computer-readable recording medium which causes a computer to
execute
the screening program according to any one of claims 7 to 13 in a case where a command stored in the recording medium is read by the computer.
15. A compound creating method of creating a three-dimensional structure of a target
compound that is bound to a target protein from a plurality of compounds, the method
comprising:
a storing step of storing a three-dimensional structure of a compound formed of a plurality of atoms and the first feature quantity calculated using the feature quantity calculating method according to claim 2 in association with each other for each of the plurality of compounds;
a feature quantity calculating step of calculating the first feature quantity of a ligand
117

that is a compound whose binding to the target protein has been confirmed using the feature quantity calculating method according to claim 2;
a generator constructing step of constructing a generator through machine learning using the three-dimensional structures of the plurality of compounds as teacher data and the first feature quantity as an explanatory variable; and
a compound three-dimensional structure generating step of generating the three-dimensional structure of the target compound from the first feature quantity of the ligand using the generator.
16. A compound creating method of creating a three-dimensional structure of a target
compound that is bound to a target protein from a plurality of compounds, the method
comprising:
a storing step of storing a three-dimensional structure of a compound formed of a plurality of atoms and the first feature quantity calculated using the feature quantity calculating method according to claim 2 in association with each other for each of the plurality of compounds;
a feature quantity calculating step of calculating the second feature quantity of the pocket structure of the target protein using the feature quantity calculating method according to claim 3;
a generator constructing step of constructing a generator through machine learning using the three-dimensional structures of the plurality of compounds as teacher data and the first feature quantity as an explanatory variable; and
a compound three-dimensional structure generating step of generating the three-dimensional structure of the target compound from the second feature quantity of the pocket structure using the generator.
17. A compound creating method of creating a three-dimensional structure of a target
compound that is bound to a target biopolymer other than a protein from a plurality of
compounds, the method comprising:
a storing step of storing a three-dimensional structure of a compound formed of a plurality of atoms and the third feature quantity in association with each other for each of the plurality of compounds;
a feature quantity calculating step of calculating the third feature quantity of a binding
118

compound that is a compound whose binding to the target biopolymer other than the protein has been confirmed using the feature quantity calculating method according to claim 4;
a generator constructing step of constructing a generator through machine learning using the three-dimensional structures of the plurality of compounds as teacher data and the third feature quantity as an explanatory variable; and
a compound three-dimensional structure generating step of generating the three-dimensional structure of the target compound from the third feature quantity of the binding compound using the generator.