Method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound and method for predicting human toxicity of a compound using the same

Abstract

The method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound according to the present invention is characterized by comprising: a data collection and preprocessing step in which a preprocessing unit collects data on the structural characteristics and toxicity status of a compound for each of inhalation toxicity, cardiac toxicity, dermal toxicity, respiratory irritation, skin irritation, and eye irritation, and organizes missing values, outliers, and duplicate data; a feature extraction step in which a feature extraction unit receives the data preprocessed in the data collection and preprocessing step and extracts features using a molecular descriptor that represents the characteristics of the compound as numbers; and a classification step in which a classifier receives the feature data extracted in the feature extraction step, learns from it, and predicts new toxicity data.

Description

Technology Field

[0001] The present invention relates to a method for predicting human toxicity of a compound, and more specifically, to a method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound and a method for predicting human toxicity of a compound using the same. Background Technology

[0003] Figure 1 is a diagram showing the method of conducting a toxicity assessment.

[0004] Referring to Figure 1, animal testing, which is currently the primary method for evaluating the toxicity of chemicals to the human body, requires significant time and cost and entails ethical issues.

[0005] Therefore, computer-based toxicity detection methods are used recently,

[0006] Using computer-based toxicity detection methods offers the advantages of cost reduction, minimization of animal use, and the ability to comprehensively consider various variables.

[0007] However, human toxicity prediction models for compounds applicable to various endpoints such as cardiotoxicity, dermal toxicity, oral toxicity, respiratory irritation, eye irritation, and skin irritation have a problem with low reliability of results. Prior art literature

[0009] KR 10-2487987B The problem to be solved

[0010] The present invention is proposed to solve the technical problems described above and provides a method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound, capable of evaluating the toxicity of a compound and a new drug regarding inhalation toxicity, cardiac toxicity, dermal toxicity, respiratory irritation, skin irritation, and eye irritation, and a method for predicting human toxicity of a compound using the same. means of solving the problem

[0012] According to one embodiment of the present invention for solving the above problems, a method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound is provided, comprising: a data collection and preprocessing step in which a preprocessing unit collects data on the structural characteristics and toxicity status of a compound for each of inhalation toxicity, cardiac toxicity, dermal toxicity, respiratory irritation, skin irritation, and eye irritation, and organizes missing values, outliers, and duplicate data; a feature extraction step in which a feature extraction unit receives the data preprocessed in the data collection and preprocessing step and extracts features using a molecular descriptor representing the characteristics of the compound as numbers; and a classification step in which a classifier receives the feature data extracted in the feature extraction step, learns, and predicts new toxicity data.

[0013] In addition, the present invention is characterized by further including a verification step of verifying the classifier with external data.

[0014] In addition, the molecular descriptor in the present invention is characterized by comprising at least one model for extracting structural features and a library for extracting physicochemical features.

[0015] In addition, the classifier in the present invention is characterized by including RF (Random Forest), XGB (XGBoost), and SVM (Support Vector Machine).

[0016] In addition, a method for predicting human toxicity of a compound is provided using a model generated through a method for generating a quantitative structure-activity relationship model for predicting human toxicity of the aforementioned compound. Effects of the invention

[0018] The method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound according to the present invention, and the method for predicting human toxicity of a compound using the same, propose a QSAR toxicity assessment model to replace animal testing. Compared to existing models, it is capable of predicting toxicity for various endpoints, including inhalation toxicity, cardiac toxicity, dermal toxicity, respiratory irritation, skin irritation, and eye irritation, and enables differentiation such as the execution of model combinations. Brief explanation of the drawing

[0020] Figure 1 is a diagram showing the method of conducting a toxicity assessment. FIG. 2 is a flowchart of the processing flowchart of the method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound of the present invention. FIG. 3 is a diagram showing the verification process of the method for generating a quantitative structure-activity relationship model for predicting human toxicity of the compound of the present invention. Figure 4 is an example of a molecular descriptor. Figure 5 is an example of a classification model (Classifier). Specific details for implementing the invention

[0021] Hereinafter, in order to explain in detail enough for a person skilled in the art to easily implement the technical concept of the present invention, embodiments of the present invention will be described with reference to the attached drawings.

[0023] FIG. 2 is a flowchart of the processing flowchart of the method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound of the present invention, FIG. 3 is a diagram showing the verification process of the method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound of the present invention, FIG. 4 is an example diagram of a molecular descriptor, and FIG. 5 is an example diagram of a classifier.

[0025] The Quantitative Structure-Activity Relationship (QSR) model is a model that predicts how the structure of a compound affects the human body.

[0027] With reference to FIGS. 2 to 5, the method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound of the present invention and the method for predicting human toxicity of a compound using the same are described as follows.

[0029]

[0030]

[0032] Table 1 is a table showing molecular descriptors that represent the characteristics of compounds in numbers.

[0033] This is a methodology for representing the characteristics of compounds numerically and is used as an input value for the model. In this invention, a total of four methods are used: a model that extracts structural features (MACCS, MORGAN, MODRED) and a model that extracts physicochemical features (Rdkit + CDK).

[0035] Table 2

[0036]

[0038] Table 2 is a table showing the classification model (Classifier).

[0039] Random Forest (RF), XGB (XGBoost), and Support Vector Machine (SVM) are used as classifiers to predict new data based on given data.

[0044] Table 3

[0045]

[0047] Table 4

[0048]

[0049] Table 3 shows the number, distribution, and source of data for the toxicity endpoint to be measured, Table 4 shows an example of visualized data, and Figure 2 is a flowchart of the processing flowchart for the method of generating a quantitative structure-activity relationship model for predicting human toxicity of the compound of the present invention.

[0051] Referring to Tables 3 and 4 and Figure 2, in the data collection and preprocessing step, the preprocessing unit collects data on the structural characteristics and toxicity of the compound for each of inhalation toxicity, cardiac toxicity, dermal toxicity, respiratory irritation, skin irritation, and eye irritation, and processes the operation of cleaning up missing values, outliers, and duplicate data.

[0053] In addition, feature extraction is performed using descriptors on the preprocessed data, and the model is trained using the extracted features as input values. At this time, three models (RF, XGB, SVM) are used for training for each descriptor.

[0054] In addition, the grid search function is used to calculate the optimal combination of parameters.

[0055] In addition, the scores for the calculated model are aggregated and saved as a .pkl file.

[0056] In addition, internal validation is performed on the trained model using 5-Fold Cross Validation to ensure that all features are validated, and the validated model is then subjected to external validation.

[0058] That is, in the feature extraction step, the feature extraction unit receives the data preprocessed in the data collection and preprocessing steps and extracts features using a molecular descriptor that represents the characteristics of the compound as a number.

[0059] In addition, in the classification stage, the classifier receives the feature data extracted in the feature extraction stage, learns from it, and predicts new toxicity data.

[0061] Figure 3 is a diagram showing the verification process of a method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound of the present invention.

[0062] In the verification stage, the above classification model (Classifier) ​​is verified using external data,

[0063] External validation is the process of using external data (test data) that was not used for training on the models generated in the aforementioned process to verify how well each model classifies the external data.

[0064] In addition, among the generated models, the optimal model can be found by employing the Grid Search method to determine which combination is best—for example, which model to use with MACCS and which model with MORGAN to achieve better results.

[0066] Table 5

[0067]

[0069] Table 5 shows a performance comparison with the existing model.

[0070] It recorded the model combination with the highest average for each final goal (Endpoint).

[0071] In the case of skin sensitization or respiratory irritation, most values ​​are low and require some improvement. Even when overall values ​​are high, there are instances where specific scores are low. (For example, if Specificity is unusually low, negative samples may be missed.)

[0072] Compared to previous studies, the average score improved by as little as 0.027 to as much as 0.07 points, and the endpoint showing the largest increase was skin irritation. However, specificity and PPV decreased.

[0074] The method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound according to the present invention, and the method for predicting human toxicity of a compound using the same, propose a QSAR toxicity assessment model to replace animal testing. Compared to existing models, it is capable of predicting toxicity for various endpoints, including inhalation toxicity, cardiac toxicity, dermal toxicity, respiratory irritation, skin irritation, and eye irritation, and enables differentiation such as the execution of model combinations.

[0076] As such, those skilled in the art to which the present invention pertains will understand that the present invention may be implemented in other specific forms without altering its technical concept or essential features. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. The scope of the present invention is defined by the claims set forth below rather than by the detailed description above, and all modifications or variations derived from the meaning and scope of the claims and their equivalents should be interpreted as being included within the scope of the present invention.

[0078] This project (result) is the result of the Local Government-University Cooperation-based Regional Innovation Project, conducted in 2024 with funding from the Ministry of Education and support from the National Research Foundation of Korea (2022RIS-006).

Claims

Claim 1 A method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound, comprising: a data collection and preprocessing step in which a preprocessing unit collects data on the structural characteristics and toxicity status of a compound for each of inhalation toxicity, cardiac toxicity, dermal toxicity, respiratory irritation, skin irritation, and eye irritation, and organizes missing values, outliers, and duplicate data; a feature extraction step in which a feature extraction unit receives the data preprocessed in the data collection and preprocessing step and extracts features using a molecular descriptor that represents the characteristics of the compound as numbers; and a classification step in which a classifier receives the feature data extracted in the feature extraction step, learns from it, and predicts new toxicity data. Claim 2 A method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound, characterized in that, in claim 1, it further includes a verification step of verifying the classifier with external data. Claim 3 A method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound, wherein, in claim 1, the molecular descriptor comprises at least one model for extracting structural features and a library for extracting physicochemical features. Claim 4 A method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound, characterized in that, in claim 1, the classifier includes RF (Random Forest), XGB (XGBoost), and SVM (Support Vector Machine). Claim 5 A method for predicting human toxicity of a compound using a model generated through a method for generating a quantitative structure-activity relationship model for predicting human toxicity of a compound of any one of claims 1 to 4.

Images