JP6941353B2 - Toxicity prediction method and its use - Google Patents

Description

The present invention relates to predicting toxicity of compounds. More specifically, it relates to methods, systems and programs for predicting the toxicity of compounds.

The toxicity of a compound is evaluated by in vitro and in vivo tests based on various toxicity indicators (eg hERG inhibition, biodegradability, mutagenicity). Unique criteria are set for each toxicity index, and the presence or absence of toxicity is determined according to the criteria. Since the test for toxicity evaluation takes a lot of time and cost, it is desirable to predict the toxicity in advance and select (narrow down) the compounds (candidate compounds) to be used in the test in advance. That is, there is a need to predict the toxicity of a compound without conducting actual tests. If toxicity can be predicted in advance, the time and cost required for the test can be reduced as the number of candidate compounds is reduced. In addition, the toxicity of compounds that cannot be actually tested or are difficult to test, such as virtual compounds, can be grasped. This advantage is particularly important in the development of new compounds, and contributes to improving the design efficiency of new compounds, improving the success rate, and reducing development costs. With the global trend of suppressing animal experiments in compound development (REACH regulation), the demand for predicting the toxicity of compounds is increasing.

In the toxicity prediction systems / programs developed so far, generally, the judgment result that the test compound has or is not toxic is output together with the prediction accuracy (for example, Patent Documents 1 to 3 and Non-Patent Documents 1 to 5). See). As the prediction accuracy, the concordance rate in cross-validation or external validation is often used. It is judged that the higher the value of the match rate, the higher the prediction accuracy. The concordance rate is defined as (the number of matches between the toxicity prediction result of the compound and the toxicity test result of the compound) / (the total number of compounds predicted to be toxic). In cross-validation, data is divided into a training set and a test set, a prediction method is constructed using the training set, and the prediction accuracy of the constructed prediction method is verified using test data. In external verification, the prediction accuracy of the constructed prediction method is verified using data independent of the data used for cross-validation. Even if cross-validation or external validation is used, there is no change in qualitative judgment (that is, with or without toxicity), and comparison between compounds (judgment of superiority or inferiority) is difficult. Moreover, since the determination is not made in relation to the structure of the compound, the reliability of the determination result cannot be said to be high.

International Publication No. 2009/025045 Pamphlet International Publication No. 2009/078096 Pamphlet International Publication No. 2010/016109 Pamphlet

Wang S., ET AL, "Recent developments in computational prediction of HERG blockage", Current Topics in Medicinal Chemistry, (The United Arab Emirates), Bentham Science Publishers, 2013, vol. 13, iss. 11, p. 1317-1326 , DOI: 10.2174 / 15680266113139990036 Blay V., ET AL, "Biodegradability Prediction of Fragrant Molecules by Molecular Topology", ACS Sustainable Chemistry and Engineering, (The United States of America), The American Chemical Society Publications, June 2016, vol. 4, iss. 8, p . 4224-4231, DOI: 10.1021 / acssuschemeng.6b00717 Jolly R., ET AL, "An evaluation of in-house and off-the-shelf in silico models: Implications on guidance for mutagenicity assessment", Regulatory Toxicology and Pharmacology, (The Netherlands), Elsevier BV, April 2015, vol. 71, iss. 3, p. 388-397, DOI: 10.1016 / j.yrtph. 2015.01.010 Greene N., ET AL, "A practical application of two in silico systems for identification of potentially mutagenic impurities", Regulatory Toxicology and Pharmacology, (The Netherlands), Elsevier BV, May 2015, vol. 72, iss. 2, p. 335-349, DOI: 10.1016 / j.yrtph.2015.05.008 Ferrari T., ET AL, "An open source multistep model to predict mutagenicity from statistical analysis and relevant structural alerts", Chemistry Central Journal, (The United Kingdom), Springer Open, July 2010, vol. 4, Suppl. 1, S2 , DOI: 10.1186 / 1752-153X-4-S1-S2 Lazar, in silico toxicology gmbh, website https://lazar.in-silico.ch/predict PASS, Vladimir Poroikov ET AL. Website http://www.pharmaexpert.ru/passonline/ HazardExpert Pro, CompuDrug International, Inc., website http://www.compudrug.com/hazardexpertpro CompuDrug International, Inc., website http://www.compudrug.com/faq

By the way, a method / system for determining the toxicity of a compound while utilizing structural information has also been developed. One of the features of the compound toxicity prediction software Lazar (Non-Patent Document 6) is that the probability of toxicity and the probability of non-toxicity are calculated and displayed for the structural information of the compound input by the user, respectively. .. However, even if the probability of toxicity and the probability of non-toxicity are added together, the probability is not 100%, and it is difficult to evaluate the prediction result. In particular, comparisons between compounds are difficult. Another compound toxicity prediction software PASS (Non-Patent Document 7) has the same problem as Laser. Software (HazardExpert Pro) (Non-Patent Document 8) has also been developed in which the sum of the probability of toxicity and the probability of non-toxicity in the prediction result is 100%. This software uses the structural information of the compound input by the user, focuses on the toxic fragment structure, calculates the probability of toxicity, and displays it. However, as CompuDrug International, Inc., which developed Hazard Expert Pro, admits that "the probability of toxicity is not an accurate value" (Non-Patent Document 9), its accuracy and reliability are not high.

Therefore, it is an object of the present invention to provide a compound toxicity prediction means capable of obtaining a prediction result having high accuracy and reliability and easy evaluation thereof.

In order to solve the above problems, the following inventions are provided.
[I] (1) A step of receiving the structural information of the test compound input by the user, and
(2) A step of generating a three-dimensional molecular structure whose structure is optimized based on the received structural information, and
(3) One or more molecular descriptions including at least one molecular descriptor selected from the group consisting of a three-dimensional molecular descriptor, a four-dimensional molecular descriptor, and a quantum chemical molecular descriptor using the three-dimensional molecular structure. Steps to generate child values and
(4) It is a step in which the toxicity prediction model calculates the probability of the presence or absence of toxicity of the test compound using the value of the molecular descriptor, and the sum of the probability of toxicity and the probability of non-toxicity is 100%. A step and
(5) A step of outputting the calculated probability and
Including
Step (4) is
(4-1) A step of normalizing the value of the molecular descriptor, and
(4-2) A step of calculating the probability of the presence or absence of toxicity of the test compound using the normalized value,
Consists of
A method of predicting the toxicity of a compound, performed by a computer.
[Ii] The three-dimensional molecular structure is a three-dimensional molecular structure whose structure is optimized by the semi-empirical molecular orbital method, a three-dimensional molecular structure whose structure is optimized by the ab initio molecular orbital method, and a density general function method. Three-dimensional molecular structure optimized by, molecular structure, semi-empirical molecular orbital method, ab initio molecular orbital method, or three-dimensional molecular structure searched by density general function method, and molecular dynamics method , A three-dimensional molecular structure whose structure is optimized by any combination of the semi-empirical molecular orbital method, the ab initio molecular orbital method, and the density general function method, with one or more three-dimensional molecular structures selected from the group consisting of. A prediction method according to [i].
[Iii] An input means for inputting structural information of the test compound, and
A receiving means for receiving the structural information of the test compound input by the user, and
A first generation means for generating a three-dimensional molecular structure whose structure is optimized based on the received structural information, and
The value of one or more molecular descriptors including at least one molecular descriptor selected from the group consisting of a three-dimensional molecular descriptor, a four-dimensional molecular descriptor, and a quantum chemical molecular descriptor using the three-dimensional molecular structure. The first calculation means for calculating
It is a calculation means for the toxicity prediction model to calculate the probability of the presence or absence of toxicity of the test compound using the value of the molecular descriptor, and when the probability of toxicity and the probability of non-toxicity are added, it is 100%. With a second calculation means, and
An output means for outputting the calculated probability and
Including
The second calculation means is a system for predicting the toxicity of a compound, which normalizes the value of the molecular descriptor and calculates the probability of the presence or absence of toxicity of the test compound using the normalized value.
[Iv] The process of receiving the structural information of the test compound input by the user, and
A process of generating a three-dimensional molecular structure with an optimized structure based on the received structural information, and
The value of one or more molecular descriptors including at least one molecular descriptor selected from the group consisting of a three-dimensional molecular descriptor, a four-dimensional molecular descriptor, and a quantum chemical molecular descriptor using the three-dimensional molecular structure. And the process of calculating
In the process in which the toxicity prediction model calculates the probability of the presence or absence of toxicity of the test compound using the value of the molecular descriptor, and the total of the probability of toxicity and the probability of non-toxicity is 100%. , The process of normalizing the value of the molecular descriptor and calculating the probability of the presence or absence of toxicity of the test compound using the normalized value, and
Processing to output the calculated probability and
A program that lets your computer run.
In addition, the present invention can also be realized in the following forms.
[1] (1) A step of receiving the structural information of the test compound input by the user, and
(2) A step of generating a three-dimensional molecular structure whose structure is optimized based on the received structural information, and
(3) One or more molecular descriptions including at least one molecular descriptor selected from the group consisting of a three-dimensional molecular descriptor, a four-dimensional molecular descriptor, and a quantum chemical molecular descriptor using the three-dimensional molecular structure. Steps to generate child values and
(4) It is a step in which the toxicity prediction model calculates the probability of the presence or absence of toxicity of the test compound using the value of the molecular descriptor, and the sum of the probability of toxicity and the probability of non-toxicity is 100%. A certain step, and (5) a step of outputting the calculated probability, and
A method of predicting the toxicity of a compound, including.
[2] The prediction method according to [1], wherein step (4) comprises the following steps.
(4-1) A step of normalizing the value of the molecular descriptor, and (4-2) a step of calculating the probability of the presence or absence of toxicity of the test compound using the normalized value.
[3] The three-dimensional molecular structure is a three-dimensional molecular structure whose structure is optimized by the semi-empirical molecular orbital method, a three-dimensional molecular structure whose structure is optimized by the ab initio molecular orbital method, and a density general function method. 3D molecular structure whose structure is optimized by , A three-dimensional molecular structure whose structure is optimized by any combination of the semi-empirical molecular orbital method, the ab initio molecular orbital method, and the density general function method, with one or more three-dimensional molecular structures selected from the group consisting of. The prediction method according to [1] or [2].
[4] The prediction method according to [1] or [2], wherein the three-dimensional molecular structure is two or more three-dimensional molecular structures whose structures have been optimized by the semi-empirical molecular orbital method.
[5] The item according to any one of [1] to [4], wherein the one or more molecular descriptors include one or more three-dimensional molecular descriptors and one or more quantum chemical molecular descriptors. Prediction method.
[6] The one or more molecular descriptors are one or more three-dimensional molecular descriptors, one or more quantum chemical molecular descriptors, one or more two-dimensional molecular descriptors, and one or more one-dimensional molecules. The prediction method according to any one of [1] to [4], which comprises a descriptor and one or more 0-dimensional molecular descriptors.
[7] Any of [1] to [6], wherein the toxicity prediction model is a toxicity prediction model constructed by machine learning using the values of normalized molecular descriptors of a plurality of compounds known to have toxicity. The prediction method described in item 1.
[8] The prediction method according to [7], wherein the machine learning is one or more machine learning selected from the group consisting of a support vector machine, a Bayesian network, a neural network, AdaBoost, a random forest, and active learning.
[9] Further including a step of generating a chemical formula of the test compound.
In step (5), the prediction method according to any one of [1] to [8], which outputs the generated chemical formula in association with the probability.
[10] The number of the test compounds is two or more,
The prediction method according to any one of [1] to [9], wherein in step (5), the probability is output for each test compound.
[11] The prediction method according to any one of [1] to [10], which outputs the determination result of the presence or absence of toxicity of the test compound together with the probability in step (5).
[12] The prediction method according to any one of [1] to [11], wherein the output of step (5) is displayed in a tabular format.
[13] The prediction method according to any one of [1] to [12], wherein the toxicity is mutagenicity determined by a return mutation test using a bacterium.
[14] An input means for inputting structural information of the test compound and
A receiving means for receiving the structural information of the test compound input by the user, and
A first generation means for generating a three-dimensional molecular structure whose structure is optimized based on the received structural information, and
The value of one or more molecular descriptors including at least one molecular descriptor selected from the group consisting of a three-dimensional molecular descriptor, a four-dimensional molecular descriptor, and a quantum chemical molecular descriptor using the three-dimensional molecular structure. The first calculation means for calculating
It is a calculation means for the toxicity prediction model to calculate the probability of the presence or absence of toxicity of the test compound using the value of the molecular descriptor, and when the probability of toxicity and the probability of non-toxicity are added, it is 100%. A second calculation means, and an output means for outputting the calculated probability.
A system for predicting the toxicity of compounds, including.
[15] An input device that functions as the input means and
An arithmetic unit that functions as the first generation means, the first calculation means, and the second calculation means.
An output device that functions as the output means,
The main memory and the control device that controls the system,
The system according to [14].
[16] The system according to [15], further comprising an auxiliary storage device in which a program is stored.
[17] The process of receiving the structural information of the test compound input by the user, and
A process of generating a three-dimensional molecular structure with an optimized structure based on the received structural information, and
The value of one or more molecular descriptors including at least one molecular descriptor selected from the group consisting of a three-dimensional molecular descriptor, a four-dimensional molecular descriptor, and a quantum chemical molecular descriptor using the three-dimensional molecular structure. And the process of calculating
A process in which the toxicity prediction model calculates the probability of the presence or absence of toxicity of the test compound using the value of the molecular descriptor, and the process is 100% when the probability of toxicity and the probability of non-toxicity are added together. , And the process of outputting the calculated probability,
A program that lets your computer run.
[18] A computer-readable storage medium containing the program according to [17].

According to the present invention, by subjecting the structural information of the compound input by the user (user) to a specific process, a prediction result with high accuracy and reliability can be output. According to the prediction result, the probability of 100% is obtained by adding the probability of toxicity and the probability of non-toxicity. Therefore, it is easy to evaluate the prediction result, that is, it is easy to compare the test compounds. For example, the present invention is carried out when a non-toxic compound is desired, and a compound having a non-toxic probability of 54% (in other words, a toxicity probability of 46%) and a non-toxic compound having a non-toxic probability of 63% (in other words, toxicity). If a compound with a probability of existence (37%) is found, the latter compound can be selected as a promising candidate by simply comparing the numerical values.

The flowchart of the toxicity prediction method of this invention. Configuration example of toxicity prediction system. An output example of the Ames mutagenicity prediction result of Example 1. Ames mutagenicity prediction result of Example 2. Structural information in SMILES format of three pesticides (anthraquinone, diquat and chlormequat) whose Ames mutagenicity was predicted in Example 3. Three-dimensional molecular structure of anthraquinone. Displayed using internal coordinates. Three-dimensional molecular structure of diquat. Displayed using internal coordinates. Three-dimensional molecular structure of chlormequat. Displayed using internal coordinates. The value of the molecular descriptor (part) used in Example 3. Ames mutagenicity prediction results of 3 pesticides (anthraquinone, diquat and chlormequat). Ames mutagenicity prediction results of Comparative Examples 1 and 2.

1. 1. Method for Predicting Toxicity of Compound The first aspect of the present invention relates to a method for predicting toxicity of a compound (hereinafter, also referred to as “prediction method of the present invention”). The prediction method of the present invention can evaluate the toxicity of a test compound (a compound whose toxicity is evaluated) without using cells or animals. If the present invention is used in combination with the toxicity evaluation using cells or animals, extremely efficient toxicity evaluation becomes possible.

In general, "toxicity of compounds" refers to acute toxicity (oral), acute toxicity (transdermal), acute toxicity (inhalation), skin corrosiveness, skin irritation, eye damage / irritation, genetic toxicity, carcinogenicity, and reproduction. It is evaluated by indicators such as toxicity, neurotoxicity, respiratory sensitivities, skin sensitivities, germ cell mutagenicity, ecotoxicity, bioaccumulation, and biodegradability. The evaluation index that defines the "toxicity of the compound" in the prediction method of the present invention is not particularly limited. In a preferred embodiment, the prediction method of the present invention is applied to the prediction of toxicity using the mutagenicity as an index determined by the bacterial reverse mutation test (Ames test) using bacteria. The Ames test method and judgment rules are stipulated in OECD TG 471.

In the present invention, the following steps (1) to (5) are performed.
(1) Step of receiving the structural information of the test compound input by the user (2) Step of generating a three-dimensional molecular structure whose structure is optimized based on the received structural information (3) The three-dimensional molecule Using the structure, the step of calculating the value of one or more molecular descriptors including at least one molecular descriptor selected from the group consisting of a three-dimensional molecular descriptor, a four-dimensional molecular descriptor, and a quantum chemical molecular descriptor. (4) It is a step in which the toxicity prediction model calculates the probability of the presence or absence of toxicity of the test compound using the value of the molecular descriptor, and the sum of the probability of toxicity and the probability of non-toxicity is 100%. A step (5) A step of outputting the calculated probability

Hereinafter, the details of the prediction method of the present invention will be described with reference to the flowchart shown in FIG. The prediction method of the present invention can be carried out by a toxicity prediction system or the like described later.

First, the structural information of the test compound input by the user (user) is received (step (1)). The user prepares the structural information of the test compound. The format of the structural information is, for example, SMILES (sometimes abbreviated as smi format) (Reference 1 and Reference 2), MDL MOL (Reference 3), SDF (Reference 3), CDX (binary file created). by PerkinElmer, Inc.'s software ChemDraw®, ChemBioDraw®), etc. The number of test compounds is one or two or more, and in the latter case, structural information is input for each test compound.

As the test compound, an organic compound having a molecular weight of 800 or less is suitable, not a polymer compound such as a protein, an antibody, a long-chain DNA or RNA, polystyrene, or a polyacrylate. Further, a compound other than a heavy metal may be used as a test compound.

Next, a three-dimensional molecular structure with an optimized structure is generated based on the received structural information (step (2)). Semi-empirical molecular orbital method, ab initio molecular orbital method, and density functional theory can be used for structural optimization, for example. Further, the structure may be optimized by conformational search by the molecular mechanics method, the semi-empirical molecular orbital method, the ab initio molecular orbital method or the density functional theory. In this step, one or more three-dimensional structures whose structures have been optimized are generated by the above optimization methods alone or in combination of two or more types. An example of optimizing the structure by using the molecular mechanics method, the semi-empirical molecular orbital method, the ab initio molecular orbital method, and the density functional theory together is shown below. In determining whether or not to use two or more types of structural optimization methods together, it is advisable to consider the length of processing time, the load applied to each device (arithmetic device, control device, main memory device, etc.), and the like.
<Example of combined use 1>
First, the structure is optimized by the semi-empirical molecular orbital method, and the obtained three-dimensional molecular structure is further optimized by the ab initio molecular orbital method or the density functional theory.
<Example 2 of combined use>
First, the structure is optimized by the Hartree-Fock method, and the resulting three-dimensional molecular structure is further optimized by the density functional theory, Moller-Plesset perturbation method, configuration interaction method, or cluster expansion method.
<Example 3 of combined use>
First, the structure is optimized by the semi-empirical molecular orbital method, the obtained three-dimensional molecular structure is further optimized by the Hartree-Fock method, and the obtained three-dimensional molecular structure is obtained by the density functional theory, Moller-. Further structural optimization is performed by the Plesset perturbation method, the configuration interaction method, or the cluster expansion method.
<Example 4 of combined use>
The structure is optimized by the Quantum Mechanics / Molecular Mechanics method or our own N-layered integrated molecular orbital and molecular mechanics method.

In general, many compounds have multiple stable three-dimensional structures depending on the conditions. Generating multiple (ie, two or more) three-dimensional structures with optimized structures reflects this point and provides more useful prediction results. In a preferred embodiment, two or more three-dimensional molecular structures whose structures are optimized by the semi-empirical molecular orbital method are generated, and the process proceeds to the next step.

"Structural optimization" is to minimize the energy of a molecule by changing the position of the atoms that make up the molecule (Reference 4). The "semi-empirical molecular orbital method" is a method of calculating the energy of the electronic state of a molecule based on an approximate expression of the Hartree-Fock equation using empirical parameters (Reference 5). The "ab initio molecular orbital method" is a method of calculating the energy of the electronic state of a molecule by the Hartree-Fock method, the Moller-Plesset perturbation method, the configuration interaction method, the cluster expansion method, or the like (Reference 6). .. The "density functional theory" is a method of calculating the energy of the electronic state of a molecule by a functional of electron density (Reference 7). The "molecular mechanics method" is a method of calculating the molecular potential energy based on the classical mechanical internuclear potential energy function (Reference 8). The "conformational search" is to systematically generate a large number of molecular conformations and then optimize the structure of each conformation (Reference 9). The Quantum Mechanics / Molecular Mechanics method and our own N-layered integrated molecular orbital and molecular mechanics method are methods for calculating the energy of a molecule by mixing (hybridizing) a plurality of the above-mentioned methods for calculating the energy of a molecule.

For the generation of structure-optimized three-dimensional molecular structures, for example, CORINA Classic (Reference 10), SYBYL®-X Suite (Reference 11), Open Babel (Reference 12), The Chemistry Development. Kit (Reference 13), RDKit (Reference 14), Chem3D ^TM (Reference 15), ChemBio3D® (Reference 16), MarvinSketch (Reference 17), Balloon (Reference 18), TINKER (Reference 18) Reference 19), Amber (Reference 20), AmberTools (Reference 21), CHARMM (Reference 22), NAMD (Reference 23), BOSS (Reference 24), VEGA ZZ / VEGA Command line (Reference 25) , GROMOS ^TM (Reference 26), GROMACS (Reference 27), MOPAC (Registered Trademark) (Reference 28), GAMESS (Reference 29), Firefly (Reference 30), Gaussian (Registered Trademark) (Reference 31) ), Spartan (Reference 32), Q-Chem (Reference 33), HyperChem (Reference 34), Molecular Operating Environment (Reference 35), BIOVIA® Discovery Studio (Reference 36), BIOVIA (Registration) Trademarks) Material Studio (Reference 37), ConfGen (Reference 38), LigPrep (Reference 39), Desmond Molecular Dynamics System (Reference 40), Jaguar (Reference 41), MacroModel (Reference 42), MOLGEN (Reference 42) Reference 43), CONFLEX® (Reference 44), OMEGA (Reference 45), VConf (Reference 46), Key3D (Reference 47), Molpro (Reference 48), Molcas (Reference 49). , ADF (reference 50), TURBOMOLE (reference 51), PQS (reference 52), MPQC (reference 53), Dalton (reference 54), LS Dalton (reference 55), COLUMBUS (reference 56), NWChem (reference 57), PSI4 (reference 58), CFOUR (reference 59), ACES (reference 60), Computer software such as ORCA (Reference 61), SMASH (Reference 62), ABINIT-MP (Reference 63), NTChem (Reference 64), PAICS (Reference 65) can be used. By instructing the above software from the outside, a three-dimensional molecular structure whose structure is optimized may be generated.

Subsequently, the values of one or more molecular descriptors are calculated using the three-dimensional molecular structure whose structure is optimized (step (3)). At least one of the molecular descriptors used is a three-dimensional molecular descriptor, a four-dimensional molecular descriptor or a quantum chemical molecular descriptor. Since the structure is based on the optimized 3D molecular structure, the value of the molecular descriptor (3D molecular descriptor, 4D molecular descriptor, quantum chemical molecular descriptor, etc.) with high accuracy or reliability is calculated. Therefore, it is possible to output highly accurate prediction results.

Preferably, two or more molecular descriptors selected from the group consisting of a three-dimensional molecular descriptor, a four-dimensional molecular descriptor, and a quantum chemical molecular descriptor (for example, a combination of a three-dimensional molecular descriptor and a four-dimensional molecular descriptor). , 3D molecular descriptor and quantum molecular descriptor combined use, 3D molecular descriptor, 4D molecular descriptor and quantum chemical molecular descriptor combined use, etc.) are calculated. For example, when a 3D molecular descriptor is used alone or in combination with another molecular descriptor (ie, a 4D molecular descriptor and / or a quantum chemical molecular descriptor), preferably 20 or more types, and further It is preferable to calculate the values of 30 or more types, and even more preferably 50 or more types of three-dimensional molecular descriptors. The upper limit of the type (number) of the three-dimensional molecular descriptor from which the value is calculated is not particularly limited. However, in consideration of the length of processing time and the load applied to each device (arithmetic unit, control device, main storage device, etc.), for example, 3,084 types can be set as the upper limit. The same applies to the four-dimensional molecular descriptor, and it is preferable to calculate values of 200 types or more, more preferably 30 types or more, and even more preferably 50 types or more (upper limit is 6,480 types, for example). The same applies to the quantum chemical molecular descriptor, and it is preferable to calculate values of 3 types or more, more preferably 5 types or more, and even more preferably 10 types or more (upper limit is 171 types, for example).

In addition to the 3D molecular descriptor, the 4D molecular descriptor and / or the quantum chemical molecular descriptor, the values of the 2D molecular descriptor, the 1D molecular descriptor, the 0-dimensional molecular descriptor, etc. are also calculated. good. Even in this case, the combination of molecular descriptors from which the value is calculated is not particularly limited. An example of a combination of molecular descriptors for which the value is calculated is as follows.
Example 1) One or more 3D molecular descriptors, 1 or more 4D molecular descriptors, 1 or more quantum chemical molecular descriptors, 1 or more 2D molecular descriptors, 1 or more 1D molecules Combination of one or more 0-dimensional molecular descriptors Example 2) One or more three-dimensional molecular descriptors, one or more quantum chemical molecular descriptors, one or more two-dimensional molecular descriptors, one or more 1-dimensional molecular descriptor of 1 or more 0-dimensional molecular descriptor Example 3) 1 or more 3-dimensional molecular descriptor, 1 or more 2-dimensional molecular descriptor, 1 or more 1-dimensional molecular descriptor Combination of one or more 0-dimensional molecular descriptors

The total number of molecular descriptors for which the value is calculated is not particularly limited, but preferably 800 or more, more preferably 1,000 or more, and even more preferably 1,500 or more molecular descriptor values are calculated. In principle, increasing the number of molecular descriptors will give more reliable prediction results. On the other hand, if the number of molecular descriptors is too large, there are adverse effects such as excessive processing time being required and excessive load being applied to each device (arithmetic unit, control device, main memory device, etc.). The total number of descriptors should be in the range of 1,000 to 10,000.

The “three-dimensional molecular descriptor” is a Radial Distribution Function descriptor, a Weighted Holistic Invariant Molecular descriptor, a Charged partial surface area, etc. (Reference 66). The "four-dimensional molecular descriptor" is Comparative Molecular Fields Analysis, GRID, conformational descriptor, four-dimensional molecular fingerprint, etc. (Reference 66 and Reference 67). The "quantum chemical molecular descriptor" is the highest occupied orbital energy, the lowest empty orbital energy, ionization potential, electron affinity, dipole moment, etc. (Reference 66 and Reference 68). The "two-dimensional molecular descriptor" is a graph conserved quantity such as Walk Count Descriptor and Path Count Descriptor, and a topology descriptor such as Topological Distance Matrix Descriptor and Zagreb Index descriptor (Reference 66). The "one-dimensional molecular descriptor" is the number of functional groups, the number of fragment structures, the molecular fingerprint, and the like (Reference 66). The "zero-dimensional molecular descriptor" is a molecular weight, the number of carbon atoms, the number of single bonds capable of free rotation, and the like (Reference 66).

For the calculation of the value of the molecular descriptor, for example, DRAGON (reference 69), CODESSA PRO (reference 70), ADAPT (reference 71), ADMET Predictor (reference 72), CORINA Symphony (reference 73), Pentacle (Reference 74), VolSurf + (Reference 75), ISIDA Fragmentor (Reference 76), JOELib (Reference 77), Molconn-Z (Reference 78), PowerMV (Reference 79), PreADMET (Reference 80) ), PaDEL-Descriptor (Reference 81), cinfony (Reference 82), Chemopy (Reference 83), The Chemistry Development Kit (Reference 13), RDKit (Reference 14), Open Babel (Reference 12), ToMoCoMD-CARDD (Reference 84), QuaSAR-Descriptor (Reference 85), Molecular Operating Environment (Reference 35), SYBYL®-X Suite (Reference 11), BIOVIA® Discovery Studio (Reference) 36), BIOVIA® Material Studio (37), QikProp (86), Jaguar (41), MacroModel (42), VCharge (87), MarvinSketch (17) ), Spartan (Reference 32), MOPAC (Registered Trademark) (Reference 28), GAMESS (Reference 29), Gaussian (Registered Trademark) (Reference 31), HyperChem (Reference 34), Q-Chem (Reference) Reference 33), BOSS (Reference 24), Firefly (Reference 30), Molpro (Reference 48), Molcas (Reference 49), ADF (Reference 50), TURBOMOLE (Reference 51), PQS (Reference 51). 52), MPQC (Reference 53), Dalton (Reference 54), LSDalton (Reference 55), COLUMBUS (Reference 56), NWChem (Reference 57), PSI4 (Reference 58), CFOUR (Reference 59) ), ACES (Reference 60), ORCA (Reference 61), SMASH (Reference 62), AB Computer software such as INIT-MP (Reference 63), NTChem (Reference 64), PAICS (Reference 65), and Mold2 (Reference 88) can be used. The value of the molecular descriptor may be calculated by instructing the above software from the outside.

Following step (3), the probability of the presence or absence of toxicity of the test compound is calculated (step (4)). A toxicity prediction model is used to calculate the probabilities. In the toxicity prediction model, the probability of the presence or absence of toxicity of the test compound is calculated using the value of the molecular descriptor calculated in step (3). One of the greatest features of the present invention is to calculate the probability of presence or absence of toxicity so that the sum of the probability of having toxicity and the probability of non-toxicity is 100%. When the number of test compounds is two or more, the probability of presence or absence of toxicity is calculated for each test compound.

As the toxicity prediction model, it is preferable to use a toxicity prediction model constructed by machine learning using the values of the normalized molecular descriptors of a plurality of compounds whose toxicity is known to be present. Examples of machine learning are support vector machines, Bayesian networks, neural networks, AdaBoost, Random Forest, and active learning. Two or more of these may be used together. For machine learning, for example, LibSVM (Reference 89), TensorFlow ^TM (Reference 90), Chainer (Registered Trademark) (Reference 91), Jubatus (Registered Trademark) (Reference 92), Caffe (Reference 93). , Theano (Reference 94), Torch (Reference 95), neon ^TM (Reference 96), MXNet (Reference 97), The Microsoft Cognitive Toolkit (Reference 98), R (C) (Reference 99), MATLAB (registered trademark) (reference 100), Mathematica (registered trademark) (reference 101), SAS (registered trademark) (reference 102), RapidMiner (registered trademark) (reference 103), KNIME (registered trademark) ( Reference 104), WeKa (Reference 105), shogun-toolbox / shogun (Reference 106), Orange (Reference 107), Apache Mahout ^TM (Reference 108), scikit-learn (Reference 1095), mlpy (Reference 104) Computer software such as Reference 110), XGBoost (Reference 111), and Deeplearning4j (Reference 112) can be used. Machine learning may be executed by instructing the above software from the outside.

Basically, the more known compounds used to build a toxicity prediction model, the more reliable the toxicity prediction model. More than 300, more preferably 3,000, and even more preferably 7,500 or more known compounds are used to build toxicity prediction models.

As step (4), the following two steps are preferably performed.
(4-1) Step of normalizing the value of the molecular descriptor (4-2) Step of calculating the probability of the presence or absence of toxicity of the test compound using the normalized value.

Step (4-1) is a step that enables comparison with the corresponding molecular descriptors of a plurality of compounds whose toxicity is known to be present. For example, calculations (processing) are performed when normalizing the values of molecular descriptors for a plurality of compounds whose toxicity is known. Using the normalized values obtained in this step, the probability of the presence or absence of toxicity of the test compound is calculated in step (4-2).

The probability calculated in step (4) is output in a predetermined format (step (5)). It is possible to output in various formats. For example, it is displayed in a table format or a graph format. Preferably, it is output so that it can be read / displayed by general-purpose software such as Excel (registered trademark) (Reference 117), Libre Office (Reference 118), and Apache Open Office ^{TM (Reference 119).} When there are two or more test compounds, the probabilities of the presence or absence of toxicity of each test compound are output, and the typical display mode is a list of the probabilities of all the test compounds. It is to be displayed, but it is not limited to this.

Along with the probability of the presence or absence of toxicity, the determination result of the presence or absence of toxicity of the test compound (typically "toxic" or "non-toxic", or something similar thereto) may be output. The determination result enables, for example, more efficiently selecting or selecting a candidate compound (a promising compound having low toxicity or expected to be non-toxic) from a plurality of test compounds.

In one aspect of the present invention, the step of generating the chemical formula of the test compound (step (6)) is also performed, and in step (5), the chemical formula generated in the step and the probability calculated in step (4) are calculated. It is output in association with each other (for example, integrated into a tabular format). Such output would indicate a link between compound structure and toxicity, providing more informative predictions. As the "chemical formula", a chemical structural formula, a molecular formula, or the like can be adopted, but preferably, the chemical structural formula is adopted from the viewpoint that the user can recognize the geometric structure of the compound. For the generation of chemical formulas, for example, The Chemistry Development Kit (Reference 13), RDKit (Reference 14), Open Babel (Reference 12), MedChem Designer ^TM (Reference 113), ChemBioDraw (Registered Trademark) (Reference). Computer software such as 114), ChemDraw® (reference 115), MarvinSketch (reference 17), BIOVIA® Draw (reference 116) can be used. The chemical formula may be generated by instructing the above software from the outside.

The prediction results according to the present invention are typically used for the next step of toxicity assessment. Specifically, based on the prediction result of the present invention, a compound to be used for toxicity evaluation using cells or animals is selected or selected. By utilizing the present invention in this way, extremely efficient toxicity evaluation is realized.

2. Toxicity Prediction System FIG. 2 is a diagram conceptually showing a configuration example of the toxicity prediction system of the present invention. The toxicity prediction system 1 of this example is a computer system including an input device 2, an arithmetic unit 3, an output device 4, a main storage device 5, a control device 6, and an auxiliary storage device 7. The solid arrow in FIG. 2 indicates the data flow direction. The broken line arrow in FIG. 2 indicates the flow direction of the control signal. The toxicity prediction system of the present invention can also be constructed using any general-purpose computer.

The input device 2 is, for example, a keyboard, a mouse, a touch panel, or the like, and the user operates the input device to input structural information of one or more test compounds. The main storage device (main memory) 5 is RAM and / or ROM. The program and data stored in the auxiliary storage device 7 are taken in and stored in the main storage device 5. The auxiliary storage device 7 is a hard disk drive, an optical disk device, an SSD, or the like. The program may be configured to be installed from a computer-readable recording medium or from another computer / server on the network or in the cloud.

The control device 6 controls other devices according to a program taken in and stored in the main storage device 5. The auxiliary storage device 7 can store the output of the computer system. The output device 4 is, for example, a display. The user can visually recognize the output of the computer system via the output device. The arithmetic unit 3 takes in the data stored in the main storage device 5, performs an operation based on the arithmetic instruction sent from the control device 6, and returns the arithmetic result to the main storage device 5.

3. 3. Programs, Storage Media The present invention also provides computer programs used in toxicity prediction systems. The computer program of the present invention causes a computer to perform the following processes (i) to (v). The computer program of the present invention includes, for example, CD (Compact Disc) -ROM, CD-R, CD-RW, DVD (Digital Versatile Disc), DVD-RAM, BD (Blu-ray (registered trademark) Disc), and the like. It is provided in a state of being stored in a computer-readable storage medium such as MO (Magneto Optical disc), SSD, magnetic tape, or various memory cards (USB flash memory, SD memory card, etc.), or downloaded from a cloud computer, etc. NS. It is also possible to store the computer program of the present invention in the auxiliary storage device of a computer connected via a network, transfer the computer program of the present invention to another computer through the network, and the like.
(I) Process of receiving the structural information of the test compound input by the user (ii) Process of generating a three-dimensional molecular structure whose structure is optimized based on the received structural information (iii) The three-dimensional molecule Process of calculating the value of one or more molecular descriptors from the structure (iv) A process of calculating the probability of the presence or absence of toxicity of the test compound using the value of the molecular descriptor by the toxicity prediction model, which is toxicity. Processing that is 100% when the probability of existence and the probability of non-toxicity are added (v) Processing that outputs the calculated probability

<Example 1: Toxicity prediction of two pesticides>
Overview, the toxicity prediction method of the present invention (FIG. 1) was carried out using the general-purpose computer system shown in FIG. To optimize the structure of the compound, we used software GAMESS, which can execute the pm3 method, which is one of the semi-empirical molecular orbital methods. In addition, from the structure-optimized 3D molecular structure, 35 types of 3D molecular descriptors, 4 types of quantum chemical molecular descriptors, 121 types of 0-dimensional molecular descriptors, 907 types of 1-dimensional molecular descriptors, and 2 types of 2-dimensional molecular descriptors. We decided to calculate 160 kinds of values. The molecular descriptor values were calculated using the software Padel-descriptor (references 81 and 121), with the exception of some molecular descriptors (calculated with reference to reference 68 and reference 120). .. On the other hand, using the values of the normalized molecular descriptors of about 8,000 compounds with known Ames mutagenicity, a toxicity prediction model was constructed by machine learning using a support vector machine.

Structural information of two pesticides (dichlobenil and teflubenzuron) in SMILES format (Clc1cccc (Cl) c1C # N, Fc1cccc (F) c1C (= O) NC (= O) Nc1cc (Cl) c (F) C (Cl) c1F) was input, the chemical structural formula of the compound and the prediction result were integrated and output in tabular format. The SMILES format structural information is the primary structure of the compound, and the format is established. (Reference 1 and Reference 2). SMILES format structure files include existing software (for example, MarvinSketch (Reference 17), ChemDraw (registered trademark) (Reference 115), BIOVIA (registered trademark) Draw (Reference 116), etc. ) Can be easily created.

FIG. 3 shows the results of predicting the mutagenicity of two pesticides as determined by a reversion mutation test using bacteria. The return mutation test using bacteria is called the Ames test. The two pesticides used as test compounds are known to have no Ames mutagenicity. As shown in FIG. 3, the probabilities of Ames mutagenicity of the two pesticides were calculated appropriately and with high accuracy. In addition, the probability of having Ames mutagenicity and the probability of not having Ames mutagenicity add up to 100%, so comparison between compounds is easy.

<Example 2: Toxicity prediction of 724 kinds of pesticides (verification of prediction accuracy)>
In order to verify the prediction accuracy of the system of the present invention, the Ames mutagenicity of 724 pesticides known to have Ames mutagenicity was predicted. As a result, the concordance rate between the prediction result by the system of the present invention and the result of the Ames test was 657/724 × 100 = 90.7 (%), which proved that the prediction accuracy of the system of the present invention was high (Fig. 4). ). In addition, the prediction system of the present invention was able to output prediction results for all 724 types. That is, it was shown to be extremely practical.

<Example 3: Comparison of prediction accuracy using three types of pesticides>
The Ames mutagenicity of three pesticides (anthraquinone, diquat and chlormequat) was predicted by the method of the present invention. The points not particularly mentioned, such as the method of optimizing the structure, are the same as those in the above embodiment.

Figure 5 shows the structural information of the three pesticides whose Ames mutagenicity was evaluated in this example in the SMILES format. In FIGS. 6 to 8, the three-dimensional molecular structures of the three pesticides generated based on the structural information are displayed using the internal coordinates. Internal coordinates consist of a bond length defined by the position of 2 atoms, a bond angle defined by the position of 3 atoms, and a helix angle defined by the position of 4 atoms. The unit of bond length is angstrom. The unit of bond angle and helix angle is ° (degrees).

Values of 35 3D molecular descriptors, 4 quantum chemical molecular descriptors, 119 0-dimensional molecular descriptors, 795 1-dimensional molecular descriptors, and 149 2D molecular descriptors based on the 3D molecular structure Was calculated. A part of the calculated value is shown in FIG. To construct the toxicity prediction model, the values of the normalized molecular descriptors of 9,719 compounds with known Ames mutagenicity were used.

The output of the prediction results of the three pesticides is shown in FIG. The probability of having Ames mutagenicity and the probability of not having Ames mutagenicity (the sum of the two probabilities is 100%) is calculated and output for each pesticide. The three pesticides used as test compounds have been found to be non-Ames mutagenic by the Ames test.

As a comparative example, the prediction result when the generated molecular descriptor was changed was obtained.
<Comparative example 1>
Values of 119 0-dimensional molecular descriptors, 795 1-dimensional molecular descriptors, and 149 2-dimensional molecular descriptors will be calculated (excluding 35 3D molecular descriptors and 4 quantum chemical molecular descriptors). ), Ames mutagenicity was predicted by the same treatment as above. The difference from the above system (Example 3) is that it does not include 35 types of 3D molecular descriptors and 4 types of quantum chemical molecular descriptors.
<Comparative example 2>
Based on the non-structure-optimized 3D molecular structure, values of 29 3D molecular descriptors, 119 0-dimensional molecular descriptors, 795 1-dimensional molecular descriptors, and 149 2D molecular descriptors are calculated. The Ames mutagenicity was predicted by the same treatment as above. The difference from the above system (Example 3) is that the value of the 3D molecular descriptor is calculated based on the 3D structure that is not structurally optimized, and does not include 4 types of quantum chemical molecular descriptors. It is a point.

The prediction results of Comparative Examples 1 and 2 are shown in FIG. Comparing with the output result of Example 3 (FIG. 10), it can be seen that in all of Comparative Examples 1 and 2, the probability of having Ames mutagenicity is high for all three types of pesticides, and the prediction accuracy and accuracy are inferior. .. Notably, in Comparative Examples 1 and 2, the probability of having Ames mutagenicity was higher than the probability of having no Ames mutagenicity for diquat, and the prediction result was contrary to the result of the Ames test.

According to the present invention, it is possible to predict the toxicity of compounds (including virtual compounds) to be developed in the future as well as known compounds. It also makes it possible to predict the toxicity of compounds for which actual testing cannot be performed or is difficult. This book is used for toxicity evaluation of chemical products, pharmaceuticals, pesticides, veterinary chemicals (livestock and pets), cosmetics, detergents, dyes, inks, additives, and other materials that require the development of non-toxic or low-toxic compounds. The invention is applicable.

The present invention is not limited to the description of the embodiments and examples of the above invention. Various modifications are also included in the present invention as long as those skilled in the art can easily conceive without departing from the description of the scope of claims. The contents of the papers, published patent gazettes, patent gazettes, etc. specified in this specification shall be cited by reference in their entirety.

<References>
1. Weininger D., "SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules", Journal of Chemical Information and Modeling, (The United States of America), The American Chemical Society Publications, February 1988, vol. 28, iss.1, p. 31-36, DOI: 10.1021 / ci00057a005
2. Weininger D., ET AL, "SMILES. 2. Algorithm for generation of unique SMILES notation", Journal of Chemical Information and Modeling, (The United States of America), The American Chemical Society Publications, May 1989, vol. 29 , iss. 2, p. 97-101, DOI: 10.1021 / ci00062a008
3. Dalby A., ET AL, "Description of several chemical structure file formats used by computer programs developed at Molecular Design Limited", Journal of Chemical Information and Computer Sciences, (The United States of America), The American Chemical Society Publications, vol. 32, iss. 3, May 1992, p. 244-255. DOI: 10.1021 / ci00007a012
4. Schlegel HB, "Geometry optimization", Wiley Interdisciplinary Reviews: Computational Molecular Science, (The United States of America), John Wiley & Sons, Ltd., May 2011, vol. 1, iss. 5, p.790-809 , DOI: 10.1002 / wcms.34
5. Zerner MC, ET AL (Eds), "Semiempirical Molecular Orbital Methods", Reviews in Computational Chemistry, (Germany), Wiley-VCH, Inc., 1991, vol. 2, p. 313-365, DOI: 10.1002 / 9780470125793.ch8
6. Friesner RA, "Ab initio quantum chemistry: Methodology and applications", Proceedings of the National Academy of Sciences of the United States of America, (The United States of America), The United States National Academy of Sciences, May 2005, vol . 102, no. 19, p. 6648-6653, DOI: 10.1073 / pnas.0408036102
7. Parr RG, "Density Functional Theory", Annual Review of Physical Chemistry, (The United States of America), Annual Reviews, Inc., 1983, vol. 34, p. 631-656, https://doi.org /10.1146/annurev.pc.34.100183.003215
8. Bowen JP ET AL, Lipkowitz KB ET AL (Eds), "Molecular Mechanics: The Art and Science of Parameterization", Reviews in Computational Chemistry, (Germany), WILEY-VCH, Inc., 1991, vol. 2, p . 81-97, DOI: 10.1002 / 9780470125793
9. Mazzanti A., ET AL, "Recent trends in conformational analysis", Wiley Interdisciplinary Reviews: Computational Molecular Science, (The United States of America), John Wiley & Sons, Inc., November 2011, vol. 2, iss. 4, p. 613-641, DOI: 10.1002 / wcms.96
10. CORINA Classic, Molecular Networks GmbH and Altamira, LLC., Website https://www.mn-am.com/products/corina
11. SYBYL®-X Suite, Certara USA, Inc., website https://www.certara.com/software/molecular-modeling-and-simulation/sybyl-x-suite/
12. O'Boyle NM, ET AL, "Open Babel: An open chemical toolbox", Journal of Cheminformatics, (The United States of America), Springer Publishing, 3:33, Oct 2011, DOI: 10.1186 / 1758-2946- 3-33
13. Steinbeck C., ET AL, "The Chemistry Development Kit (CDK): An Open-Source Java Library for Chemo- and Bioinformatics", Journal of Chemical Information and Computer Sciences, (The United States of America), The American Chemical Society Publications, vol. 43, iss. 2, February 2003, p. 493-500, DOI: 10.1021 / ci025584y
14. RDKit, Landrum G., "RDKit Documentation Release 2017.03.1", Online Documentation, Website http://www.rdkit.org/docs/Overview.html
15. Chem3D ^TM , PerkinElmer, Inc., website http://www.cambridgesoft.com/Ensemble_for_Chemistry/ChemOffice/ChemOfficeProfessional/
16. ChemBio3D®, PerkinElmer, Inc., website https://www.cambridgesoft.com/Ensemble_for_Chemistry/details/Default.aspx?fid=17&pid=660
17. MarvinSketch, ChemAxon, website https://www.chemaxon.com/products/calculator-plugins/molecular-modelling/
18. Balloon, Vainio M., website http://users.abo.fi/mivainio/balloon/
19. TINKER, Ponter J., ET AL, Website https://dasher.wustl.edu/tinker/
20. Amber, Case DA, ET AL, Website http://ambermd.org/
21. AmberTools, Website http://ambermd.org/#AmberTools
22. CHARMM, Karplus M., ET AL, Website https://www.charmm.org/charmm/
23. NAMD, Theoretical and Computational Biophysics Group, Beckman Institute, University of Illinois, Website http://www.ks.uiuc.edu/Research/namd/
24. BOSS, Jorgensen WL, ET AL, Website http://zarbi.chem.yale.edu/software.html
25. VEGA ZZ / VEGA Command line, Pedretti A., ET AL: Website http://www.vegazz.net/
26. GROMOS ^TM , van Gunsteren WF, ET AL, website http://www.igc.ethz.ch/gromos.html
27. GROMACS, van der Spoel D., ET AL, website http://www.gromacs.org/
Website http://openmopac.net/
29. GAMESS, Schmidt MW, ET AL, website http://www.msg.ameslab.gov/gamess/
30. Firefly, Granovsky AA, ET AL, website http://classic.chem.msu.su/gran/gamess/index.html
31. Gaussian®, Gaussian, Inc., website http://gaussian.com/citation/
32. Spartan, Wavefunction, Inc., Website https://www.wavefun.com/products/spartan.html
33. Q-Chem, Q-Chem, Inc., website http://www.q-chem.com/
34. HyperChem, Hypercube, Inc., website http://www.hyper.com/
35. Molecular Operating Environment, Chemical Computing Group ULC, Website https://www.chemcomp.com/MOE-Molecular_Operating_Environment.htm
36. BIOVIA® Discovery Studio, Dassault Systemes, website http://accelrys.com/products/collaborative-science/biovia-discovery-studio/
37. BIOVIA® Material Studio, Dassault Systemes, website http://accelrys.com/products/collaborative-science/biovia-materials-studio/
38. ConfGen, Schrodinger®, LLC, Schrodinger Release 2017-1: New York, NY, 2017 Website https://www.schrodinger.com/confgen
39. LigPrep, Schrodinger®, LLC, Schrodinger Release 2017-1: New York, NY, 2017, Website https://www.schrodinger.com/ligprep
40. Desmond Molecular Dynamics System, DE Shaw Research, New York, NY, 2017. Website https://www.schrodinger.com/desmond
41. Jaguar, Schrodinger®, LLC, Schrodinger Release 2017-1: New York, NY, 2017, Website https://www.schrodinger.com/jaguar
42. MacroModel, Schrodinger®, LLC, Schrodinger Release 2017-1: New York, NY, 2017, Website https://www.schrodinger.com/macromodel
43. MOLGEN, Wassermann A., ET AL, website http://www.molgen.de/
44. CONFLEX®, CONFLEX Corporation, website http://www.conflex.net/
45. OMEGA, OpenEye Scientific Software, Website https://www.eyesopen.com/omega
46. VConf, VeraChem, LLC, website http://www.verachem.com/products/vconf/
47. Key3D, IMMD, Inc., website http://www.immd.co.jp/en/product_2.html
48. Molpro, TTI GmbH, Website https://www.molpro.net/
49. Molcas, Veryazov. V. ET AL, website http://www.molcas.org/
50. ADF®, Scientific Computing & Modeling NV, Website https://www.scm.com/product/adf/
51. TURBOMOLE, TURBOMOLE GmbH, website http://www.turbomole.com/
52. PQS, Parallel Quantum Solutions, website http://www.pqs-chem.com/
53. MPQC, Valeev E. ET AL, Website http://www.mpqc.org/
54. Dalton, Dalton / LS Dalton developers, website http://daltonprogram.org
55. LSDalton, Dalton / LS Dalton developers website http://daltonprogram.org
56. COLUMBUS, Lischka H. ET AL, website https://www.univie.ac.at/columbus/
57. NWChem, Valiev M., ET AL, Website http://www.nwchem-sw.org/index.php/Main_Page
58. PSI4, Sherrill CD, ET AL, website http://www.psicode.org/
59. CFOUR, Stanton JF, ET AL, website http://www.cfour.de/
60. ACES, Bartlett RJ, ET AL, website http://www.qtp.ufl.edu/ACES/
61. ORCA, Neese, F., ET AL, website https://orcaforum.cec.mpg.de/
62. SMASH, Ishimura K., Website http://smash-qc.sourceforge.net/
63. ABINIT-MP®, Mochizuki Y., ET AL, Website http://www.cenav.org/abinitmpopen1/
64. NTChem, Nakajima, T., ET AL, Website http://labs.aics.riken.jp/nakajimat_top/ntchem_e.html
65. PAICS, Ishikawa T., Website http://www.paics.net/index_e.html
66. Johann Gasteiger J., ET AL (Editor), "Chemoinformatics: A Textbook", WILEY-VCH Verlag GmbH & Co. KgaA, Weinheim, 2003, ISBN 3-527-30681-1
67. Wicker JGP, ET AL, "Beyond Rotatable Bond Counts: Capturing 3D Conformational Flexibility in a Single Descriptor", Journal of Chemical Information and Modeling, (The United States of America), The American Chemical Society Publications, vol. 56, iss . 12, November 2016, p. 2347-2352, DOI: 10.1021 / acs.jcim.6b00565
68. Karelson M., ET AL, "Quantum-Chemical Descriptors in QSAR / QSPR Studies", Chemical Reviews, (The United States of America), The American Chemical Society Publications, vol. 96, iss. 3, May 1996, p. . 1027-1044, DOI: 10.1021 / cr950202r
69. DRAGON, Kode srl, website https://chm.kode-solutions.net/products_dragon.php
70. CODESSA PRO, CompuDrug International, Inc., website http://www.compudrug.com/codessa_pro
71. ADAPT, Jurs PC, ET AL, website http://research.chem.psu.edu/pcjgroup/adapt.html
72. ADMET Predictor, Simulations Plus, Inc., website http://www.simulations-plus.com/software/admet-property-prediction-qsar/
73. CORINA Symphony, Molecular Networks GmbH and Altamira, LLC., Website https://www.mn-am.com/products/corinasymphony
74. Pentacle, Molecular Discovery Ltd, website http://www.moldiscovery.com/software/pentacle/
75. VolSurf +, Molecular Discovery Ltd, website http://www.moldiscovery.com/software/vsplus/
76. ISIDA Fragmentor, Varnek A., ET AL, website http://infochim.u-strasbg.fr/spip.php?rubrique41
77. JOELib, Zell A., ET AL, Website http://www.ra.cs.uni-tuebingen.de/software/joelib/index.html
78. Molconn-Z, eduSoft, LC, website http://www.edusoft-lc.com/molconn/
79. PowerMV, Liu J., ET AL, Website https://www.niss.org/research/software/powermv
80. PreADMET, Bioinformatics & Molecular Design Research Center (BMDRC), Website https://preadmet.bmdrc.kr/preadmet-pc-version-2-0/
81. PaDEL-Descriptor, Yap CW, Website http://www.yapcwsoft.com/dd/padeldescriptor/
82. Cinfony, O'Boyle NM, ET AL, website http://cinfony.github.io/
83. Cao D.-S., ET AL, "ChemoPy: freely available python package for computational biology and chemoinformatics", Bioinformatics, (The United Kingdom), Oxford University Press, vol. 29, iss. 8, March 2013, p . 1092-1094, DOI: https://doi.org/10.1093/bioinformatics/btt105
84. ToMoCoMD-CARDD, Ponce YM, Website http://tomocomd.com/
85. QuaSAR-Descriptor, Chemical Computing Group ULC, website https://www.chemcomp.com/journal/descr.htm
86. QikProp, Schrodinger®, LLC, Schrodinger Release 2017-1: New York, NY, 2017, Website https://www.schrodinger.com/qikprop
87. VCharge, VeraChem, LLC, website http://www.verachem.com/products/vcharge/
88. Mold2, Hong, H., ET AL, Website https://www.fda.gov/ScienceResearch/BioinformaticsTools/Mold2/ucm144528.htm
89. LibSVM, Chang C.-C., ET AL, Website https://www.csie.ntu.edu.tw/~cjlin/libsvm/
90. TensorFlow ^TM , Google Inc., Website https://www.tensorflow.org/
91. Chainer®, Preferred Networks, Inc., Website http://chainer.org/
92. Jubatus®, Preferred Networks, Inc. and Nippon Telegraph and Telephone Corporation, website https://jubat.us/en/
93. Caffe, Jia Y., ET AL, website http://caffe.berkeleyvision.org/
94. Theano, Theano Development Team, Website http://deeplearning.net/software/theano/
95. Torch, Ronan Collobert ET AL, website http://torch.ch/
96. neon ^TM , Nervana Systems, website https://www.nervanasys.com/technology/neon/
97. MXNet, website http://mxnet.io/
98. The Microsoft Cognitive Toolkit, Microsoft Corporation, Website https://www.microsoft.com/en-us/cognitive-toolkit/
99. R (c), The R Foundation, Website https://www.r-project.org/
100. MATLAB®, The MathWorks, Inc., Website https://www.mathworks.com/products/matlab.html?s_tid=hp_products_matlab
101. Mathematica®, Wolfram Research, Website http://www.wolfram.com/mathematica/
102. SAS®, SAS Institute Inc., website https://www.sas.com/en_us/home.html
103. RapidMiner®, RapidMiner, Inc., website https://rapidminer.com/
104. KNIME®, KNIME.com AG, website https://www.knime.org/
105. Witten IH, ET AL, The WEKA Workbench. Online Appendix for "Data Mining: Practical Machine Learning Tools and Techniques", (The United States of America), Morgan Kaufmann, Fourth Edition, 2016, ISBN13: 978-0128042915. Web Site http://www.cs.waikato.ac.nz/ml/weka/
106. shogun-toolbox / shogun, website http://www.shogun-toolbox.org/, Sonnenburg S., ET AL, "shogun-toolbox / shogun: Shogun 6.0.0 --Baba Nobuharu", April 2017, DOI : 10.5281 / zenodo.556748
107. Orange, Demsar J., ET AL, "Orange: Data Mining Toolbox in Python", Journal of Machine Learning Research, (The United States of America), JMLR, Inc. and Microtome Publishing, vol. 14, Aug 2013, p. 2349-2353. Website https://orange.biolab.si/
108. Apache Mahout ^TM , The Apache Software Foundation, Website http://mahout.apache.org/
109. scikit-learn, website http://scikit-learn.org/stable/. Pedregosa F., ET AL, "Scikit-learn: Machine Learning in Python", Journal of Machine Learning Research, (The United States of America), JMLR, Inc. and Microtome Publishing, vol. 12, Oct 2011, p. 2825-2830
110. mlpy, Albanese D., ET AL, website http://mlpy.sourceforge.net/
111. Chen T., ET AL, "XGBoost: A Scalable Tree Boosting System", Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Aug 2016, DOI: 10.1145 / 2939672.2939785
112. Deeplearning4j, Skymind, Deeplearning4j Development Team, "Deeplearning4j: Open-source distributed deep learning for the JVM", Apache Software Foundation License 2.0. Website https://deeplearning4j.org/
113. MedChem Designer ^TM , Simulations Plus, Inc., Website http://www.simulations-plus.com/software/medchem-designer/
114. ChemBioDraw®, PerkinElmer, Inc., website https://www.cambridgesoft.com/Ensemble_for_Chemistry/details/Default.aspx?fid=17&pid=660
115. ChemDraw®, PerkinElmer, Inc., website http://www.cambridgesoft.com/software/overview.aspx
116. BIOVIA® Draw, Dassault Systemes, website http://accelrys.com/products/collaborative-science/biovia-draw/
117. Excel®, Microsoft Corporation, website https://products.office.com/en-us/excel
118. Libre Office, The Document Foundation, Website https://www.libreoffice.org/
119. Apache Open Office ^TM , The Apache Software Foundation, Website https://www.openoffice.org/
120. Sakuratani Y, ET AL, "Molecular size as a limiting characteristic for bioconcentration in fish", Journal of Environmental Biology, January 2008, vol. 29, iss. 1, p. 89-92., Website http: // www.jeb.co.in/journal_issues/200801_jan08/paper_15.pdf
121. Yap CW, "PaDEL-descriptor: An open source software to calculate molecular descriptors and fingerprints", Journal of Computational Chemistry, Volume 32, Issue 7, p. 1466-1474, May 2011. DOI: 10.1002 / jcc.21707. Supporting Information: JCCT_21707_sm_suppinformation.xls, website http://onlinelibrary.wiley.com/doi/10.1002/jcc.21707/suppinfo

1 Toxicity prediction system 2 Input device 3 Arithmetic device 4 Output device 5 Main memory device 6 Control device 7 Auxiliary storage device

Claims (17)

(1) A step of receiving the structural information of the test compound input by the user, and
(2) A step of generating a three-dimensional molecular structure whose structure is optimized based on the received structural information, and
(3) One or more molecular descriptions including at least one molecular descriptor selected from the group consisting of a three-dimensional molecular descriptor, a four-dimensional molecular descriptor, and a quantum chemical molecular descriptor using the three-dimensional molecular structure. Steps to generate child values and
(4) It is a step in which the toxicity prediction model calculates the probability of the presence or absence of toxicity of the test compound using the value of the molecular descriptor, and the sum of the probability of toxicity and the probability of non-toxicity is 100%. A certain step, and (5) a step of outputting the calculated probability, and
Only including,
Step (4) is
(4-1) A step of normalizing the value of the molecular descriptor, and
(4-2) A step of calculating the probability of the presence or absence of toxicity of the test compound using the normalized value,
Consists of
A method of predicting the toxicity of a compound, performed by a computer. The three-dimensional molecular structure is the three-dimensional molecular structure whose structure is optimized by the semi-empirical molecular orbital method, the three-dimensional molecular structure whose structure is optimized by the ab initio molecular orbital method, and the structure by the density general function method. Optimized 3D molecular structure , molecular dynamics method, semi-empirical molecular orbital method, ab initio molecular orbital method or 3D molecular structure searched by density general function method, and molecular dynamics method, semi-experience One or more three-dimensional molecular structures selected from the group consisting of a three-dimensional molecular structure whose structure is optimized by any combination of the ab initio molecular orbital method, the ab initio molecular orbital method, and the density general function method. Item 1. The prediction method according to Item 1. The prediction method according to claim 1 , wherein the three-dimensional molecular structure is two or more three-dimensional molecular structures whose structures have been optimized by the semi-empirical molecular orbital method. The prediction method according to any one of claims 1 to 3 , wherein the one or more molecular descriptors include one or more three-dimensional molecular descriptors and one or more quantum chemical molecular descriptors. The one or more molecular descriptors are one or more three-dimensional molecular descriptors, one or more quantum chemical molecular descriptors, one or more two-dimensional molecular descriptors, and one or more one-dimensional molecular descriptors. The prediction method according to any one of claims 1 to 3 , which comprises one or more 0-dimensional molecular descriptors. The toxicity prediction model according to any one of claims 1 to 5 , wherein the toxicity prediction model is a toxicity prediction model constructed by machine learning using the values of normalized molecular descriptors of a plurality of compounds known to have toxicity. Prediction method. The prediction method according to claim 6 , wherein the machine learning is one or more machine learning selected from the group consisting of a support vector machine, a Bayesian network, a neural network, AdaBoost, a random forest, and active learning. It further comprises the step of generating the chemical formula of the test compound.
The prediction method according to any one of claims 1 to 7 , wherein in step (5), the generated chemical formula is associated with the probability and output. There are two or more of the test compounds,
The prediction method according to any one of claims 1 to 8 , wherein in step (5), the probability is output for each test compound. The prediction method according to any one of claims 1 to 9 , wherein in step (5), the determination result of the presence or absence of toxicity of the test compound is output together with the probability. The prediction method according to any one of claims 1 to 10 , wherein the output of step (5) is displayed in a tabular format. The prediction method according to any one of claims 1 to 11, wherein the toxicity is mutagenicity determined by a return mutation test using a bacterium. An input means for inputting the structural information of the test compound,
A receiving means for receiving the structural information of the test compound input by the user, and
A first generation means for generating a three-dimensional molecular structure whose structure is optimized based on the received structural information, and
The value of one or more molecular descriptors including at least one molecular descriptor selected from the group consisting of a three-dimensional molecular descriptor, a four-dimensional molecular descriptor, and a quantum chemical molecular descriptor using the three-dimensional molecular structure. The first calculation means for calculating
It is a calculation means for the toxicity prediction model to calculate the probability of the presence or absence of toxicity of the test compound using the value of the molecular descriptor, and when the probability of toxicity and the probability of non-toxicity are added, it is 100%. A second calculation means, and an output means for outputting the calculated probability.
Only including,
The second calculation means is a system for predicting the toxicity of a compound, which normalizes the value of the molecular descriptor and calculates the probability of the presence or absence of toxicity of the test compound using the normalized value. An input device that functions as the input means and
An arithmetic unit that functions as the first generation means, the first calculation means, and the second calculation means.
An output device that functions as the output means,
The main memory and the control device that controls the system,
Containing system of claim 1 3. Further comprising an auxiliary storage device in which the program is stored, the system according to claim 1 4. The process of receiving the structural information of the test compound input by the user, and
A process of generating a three-dimensional molecular structure with an optimized structure based on the received structural information, and
The value of one or more molecular descriptors including at least one molecular descriptor selected from the group consisting of a three-dimensional molecular descriptor, a four-dimensional molecular descriptor, and a quantum chemical molecular descriptor using the three-dimensional molecular structure. And the process of calculating
In the process in which the toxicity prediction model calculates the probability of the presence or absence of toxicity of the test compound using the value of the molecular descriptor, and the total of the probability of toxicity and the probability of non-toxicity is 100% . , A process of normalizing the value of the molecular descriptor and calculating the probability of the presence or absence of toxicity of the test compound using the normalized value , and a process of outputting the calculated probability.
A program that lets your computer run. A computer-readable storage medium containing the program according to claim 16.

Images