KR20200072585A - Method for predicting the HAZARD and RISK of target chemicals BASED ON AI - Google Patents

Abstract

According to an embodiment of the present invention, a method for predicting harmfulness and risk of a target substance based on artificial intelligence, which is performed by a physical property prediction device for predicting physical properties of a target substance, comprises: a pattern analysis step of analyzing a character string pattern for chemical properties or a structure of the target substance; a prediction modeling step of providing a physical property analysis and a harmfulness and risk prediction model based on artificial intelligence by using data about physical properties and properties of at least one chemical substance as learning data, and calculating a prediction value according to similarity for the analyzed character string pattern based on a toxicity prediction model; and an information providing step of comparing the calculated prediction value with an experimental value of an experimental substance having chemical properties similar to those of the target substance so as to predict physical properties, harmfulness and risk of the target substance and determine whether the target substance meets the chemical substance regulatory standards, and matching the determined physical properties, harmfulness and risk with preset chemical substance regulatory standards to provide physical properties and toxicity determination result including a matching result.

Description

Method for predicting the HAZARD and RISK of target chemicals BASED AI based on artificial intelligence

The present invention relates to a method for predicting the harmfulness and risk of a target substance based on artificial intelligence for predicting the physical properties of the target substances and a possibility of causing harm to the human body and the environment by combining a plurality of analysis models based on artificial intelligence.

In general, when a new chemical substance is developed, the developer of the new substance undergoes a post-test to determine whether it conforms to the regulations on the registration and evaluation of chemical substances (hereinafter referred to as the Peace Act) and the specific methods of chemical risk assessment, etc. When the documents are submitted, the government-related departments grade new substances by passing, restricting, or prohibiting them according to the guidelines of the Peace Act.

Here, the major evaluation areas of the Peace Act are physical and chemical properties, human hazards, and environmental hazards.In the business field to which the human hazard assessment is applied, the pharmaceutical and medical fields are bio, and to the business field to which the environmental hazard assessment is applied, chemical and It is a petrochemical/semiconductor and electronic/energy company, and is a wastewater and industrial waste sector.

At this time, new material developers can predict the physical properties including approximate toxicity in chemical formulas when designing new substances in advance, but the accuracy is not high, and in the case of complex and complex chemical substances, it is difficult to grasp the physical properties only with chemical formula information. The developer performs post-tests.

In the post-testing process for new substances, there is a problem in that the test itself is expensive, and there is a problem that the experimenter is exposed to toxicity due to the exposure of toxic substances, so that safety cannot be guaranteed. If it is determined as, there is a problem that the costs such as raw materials and labor costs for the development of the material become a sunk cost.

As the voice of concern about the harmfulness of chemicals to the human body and the environment increases, the peace law is strengthening, and the development of new materials is also increasing.

In the field of chemical development, Cheminformatics, which analyzes big data to predict physical properties, is emerging, and a leading company developed a physical property predictor overseas, and some domestic pharmaceutical companies introduced a physical property predictor. There is a problem that utilization is limited to the pharmaceutical field, such as predicting mutagenicity or predicting carcinogenesis.

Recently, as concerns about humidifier disinfectants, sanitary napkin surges, and environmental pollution of industrial wastes have increased, the justification for the mandatory inspection of chemical substances has increased, and as a result, the criteria for evaluation of new chemical substances by the Peace Act have become stricter At the same time, the process has also been complicated.

In the case of a new new substance or a substance registered under the first peace law, it can cause cost or time disadvantages in the development of new substances because it can act as a failure factor in registration in confirming the hazards and risks during registration.

Therefore, a new material developer or developer needs a technology that can check the environmental impact in advance before designing the chemical structure of a new material, and apply the results to the production process quickly and predictive analysis to keep up with trends.

The present invention, in order to solve the above-mentioned problems, according to an embodiment of the present invention, by combining a number of analytical models capable of analyzing and predicting properties and toxicity for new substances, artificial intelligence capable of predicting the possibility of causing toxicity of target substances It is an object to provide a method for predicting the hazard and risk of a target substance based on.

However, the technical problem to be achieved by the present embodiment is not limited to the technical problem as described above, and other technical problems may exist.

As a technical means for achieving the above technical problem, a method for predicting the harmfulness and risk of a target material based on deep learning performed by a property prediction device for predicting a property of the target material according to an embodiment of the present invention includes the target material A pattern analysis step of analyzing a string pattern for a chemical property or structure of the; Provides a prediction model based on deep learning using data on the properties and properties of at least one chemical substance as learning data, and analyzes the string pattern and molecular structure characteristics (geometrical stereoscopic, electronic based on the prediction model based on the toxicity prediction model) Predictive modeling step of calculating a predicted value according to the similarity to characteristics; And comparing the calculated predicted value with an experimental value of a test substance having chemical characteristics similar to the target substance, to determine the toxicity of the target substance, and matching the determined toxicity with a preset chemical substance regulation criterion to include properties of the result including matching results. And providing an information providing step of providing a toxicity determination result.

In the predictive modeling step, at least one analysis model is combined to group chemical properties having a predetermined similarity range, and classification and regression analysis is performed by determining the contribution to the string pattern of the target material from the clustered chemical properties. Is to do.

Here, the at least one analysis model, Scikit-learn, Keras, Tensorflow, PyTorch, k-N (k-Nearest Neighbors), decision tree, Random Forest (RF), gradient boosting, ensemble, SVM (Suppoert vector machine) , Naive Bayes, Regression, K-means, PCA (principal component analysis), clustering, Grid Search, Cross Validation, ANN, CNN, RNN, Autoencoder, GAN, DNN, QNN, ANOVA (Analysis of variance) To include.

On the other hand, in the pattern analysis step, the chemical structure of the target substance is converted into a one-dimensional character array and then compared using SMILES (Simplified molecular-input line-entry system) code and various naming notations representing chemical structural formulas. It is to analyze the similarity of patterns and molecular structures.

In the step of providing the information, the half-lethal concentration (LC50) or the half-water impact concentration (EC50) of the subject is collected, and the harmfulness of the target material is based on the collected half-lethal concentration or half-water impact concentration, and other physicochemical properties It is a numerical prediction of the risks and the determination of compliance with chemical regulations.

In the providing of the information, a correlation between a chemical structure and a hazard of the target substance is analyzed through a static analysis, and a dose-response curve and a predicted value result for the relationship between the chemical property or structure of the target substance and environmental hazard and risk, and the It is to provide legality as a result of toxicity determination as to whether or not it is compatible with a predetermined chemical substance regulation standard. The predetermined chemical substance regulation standards consist of evaluation criteria items for evaluating the target substance as a licensed substance in accordance with the Act on Registration and Evaluation of Chemical Substances.

The step of providing information is to redesign and provide a chemical structure that reduces toxicity of the target substance in association with the predicted value.

According to the above-described problem solving means of the present invention, it is possible to quickly and easily perform a pattern analysis by converting a molecular formula of a new material into a character string, and performing a character string analysis using a number of deep learning algorithms to achieve the accuracy of'pattern analysis' and 'Correlation of chemical structure with hazards and risks' and'pattern diversity' can be secured.

Since the present invention can predict the toxic properties of the target substance prior to the actual analysis by understanding the relationship between the chemical structural formula and the toxicity of the target substance, it ensures safety in terms of environmental protection when developing a new substance, and improves the efficiency of smoke by reducing time and cost. Can. It is possible to determine whether the target substance meets the environmental hazard standards.

In addition, the present invention modularizes the pattern analysis process or the predictive modeling process, and through this module, the application and understanding of chemical experiments can be improved, existing chemicals can be used as learning data, and toxicity of new materials can be determined. The results can be used as consulting data when registering the Peace Act.

1 is a view for explaining a configuration of a property prediction device for implementing a method for predicting the harmfulness of a target material based on deep learning according to an embodiment of the present invention.
FIG. 2 is a diagram showing a detailed configuration of the processor of FIG. 1.
3 is an operation flowchart illustrating a method for predicting the harmfulness of a target material based on deep learning according to an embodiment of the present invention.
4 is an exemplary view illustrating a chemical structure of a target material input in a pattern analysis step according to an embodiment of the present invention.
5 is an exemplary diagram illustrating a pattern analysis state for a target material in a pattern analysis step and a predictive modeling step according to an embodiment of the present invention.
6 is a view for explaining a process of visualizing and outputting a result of toxic discrimination in an information providing step according to an embodiment of the present invention.
7 and 8 are diagrams for explaining a prediction model of toxicity based on a convolutional neural network applied to a method for predicting harmfulness and risk of a target substance based on artificial intelligence.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art to which the present invention pertains may easily practice. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. In addition, in order to clearly describe the present invention in the drawings, parts irrelevant to the description are omitted, and like reference numerals are assigned to similar parts throughout the specification.

Throughout the specification, when a part is "connected" to another part, this includes not only "directly connected" but also "electrically connected" with another element in between. . Also, when a part is said to "include" a certain component, it means that the component may further include other components, not exclude other components, unless specifically stated otherwise. However, it should be understood that the existence or addition possibilities of numbers, steps, actions, components, parts or combinations thereof are not excluded in advance.

The following examples are detailed descriptions to aid the understanding of the present invention, and do not limit the scope of the present invention. Therefore, the same scope of the invention performing the same function as the present invention will also belong to the scope of the present invention.

Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

1 is a view for explaining the configuration of a property prediction device for implementing a method for predicting the harmfulness of a target material based on artificial intelligence according to an embodiment of the present invention.

Referring to FIG. 1, the property prediction apparatus 100 includes a communication module 110, a memory 120, a processor 130, and a database 140.

In detail, the communication module 110 provides a communication interface required to provide a transmission/reception signal in the form of packet data to the property prediction apparatus 100 in conjunction with the communication network 300. Here, the communication module 110 may be a device including hardware and software necessary for transmitting and receiving a signal such as a control signal or a data signal through a wired or wireless connection with another network device.

In the memory 120, a program for performing a method for predicting the harmfulness of a target substance based on artificial intelligence is recorded. In addition, it performs a function of temporarily or permanently storing data processed by the processor 130. Here, the memory 120 may include volatile storage media or non-volatile storage media, but the scope of the present invention is not limited thereto.

The processor 130 controls the entire process of providing a method for predicting the harmfulness of a target material based on artificial intelligence. Each step performed by the processor 130 will be described later with reference to FIGS. 2 and 3.

Here, the processor 130 may include all kinds of devices capable of processing data. Here, a'processor' may mean a data processing device embedded in hardware having physically structured circuits, for example, to perform functions represented by codes or instructions included in a program. As an example of such a data processing device embedded in hardware, a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, and an application-specific integrated ASIC circuit), a field programmable gate array (FPGA), and the like, but the scope of the present invention is not limited thereto.

The database 140 stores data accumulated while performing a method of predicting the harmfulness of a target substance based on artificial intelligence. For example, the database 140 may store learning data, chemical properties and structures of the target material, pattern analysis results, and toxicity determination results.

FIG. 2 is a diagram showing a detailed configuration of the processor of FIG. 1.

Referring to FIG. 2, the processor 130 includes, but is not limited to, a pattern analysis module 131, a prediction modeling module 132, and an information providing module 133.

The pattern analysis module 131 analyzes a character string pattern for a chemical property or structure of a target material.

The predictive modeling module 132 provides a toxicity prediction model based on artificial intelligence, and calculates a predicted value of whether the target substance is toxic based on the toxicity prediction model.

The information providing module 133 determines the toxicity of the target substance, and visualizes the property and toxicity determination result including the matching result by providing the user with the determined toxicity by matching the predetermined chemical regulation standard.

As such, the processor 130 individually reinforces and learns the pattern analysis module 131, the prediction modeling module 132, and the information providing module 133 to improve performance for each module.

3 is an operation flowchart showing a method for predicting the harmfulness of a target material based on artificial intelligence according to an embodiment of the present invention, and FIG. 4 shows the chemical structure of the target material input in the pattern analysis step according to an embodiment of the present invention 5 is an exemplary diagram illustrating a pattern analysis state for a target material in a pattern analysis step and a predictive modeling step according to an embodiment of the present invention, and FIG. 6 is a diagram according to an embodiment of the present invention It is a diagram explaining the process of visualizing and outputting the result of toxicity determination in the information provision step.

Referring to FIGS. 3 to 6, in a method for predicting the harmfulness of a target substance based on artificial intelligence, when the chemical formula of the target substance is input as a SMILES (Simplified molecular-input line-entry system) code (S110), the chemical properties of the target substance Alternatively, character string patterns and molecular structure characteristics (geometry, electronic characteristics, etc.) of the structure are analyzed (S120). In addition to the SMILES code, other nomenclature and molecular structure diagrams can also be entered.

As a method of classifying chemical substances, there is a conversion method in which a three-dimensional structure of a chemical substance is converted into a one-dimensional character array and then compared. The most commonly used methods for this conversion method are SMILES and InChI (International Chemical Identifier). The character array generated by this conversion method has the advantage of reducing the size of the DB than the method of storing chemical substances in a three-dimensional structure.

In particular, SMILES linearly displays atoms contained in a molecule as shown in FIG. 4 to indicate the arrangement and bonding of atoms. Therefore, the three-dimensional structure represented by SMILES has a merit to represent a simple three-dimensional structure with good readability by quickly and elaborately analyzing complex and long molecular formulas.

As shown in FIG. 5, the processor analyzes the chemical structure pattern of chemical formulas named using RNN, CNN, GRU, and GAN algorithms to analyze string patterns, extracts features of chemical molecular formula structures, processes text, and chemical structure patterns Improve analysis accuracy, design new chemical structures, and more.

Meanwhile, the processor 130 provides a toxicity prediction model based on deep learning by using data on physical properties, physical or chemical properties, and biological information of at least one chemical substance as training data, and is analyzed based on the toxicity prediction model The predicted value is calculated according to the similarity to the string pattern (S130). Here, the toxicity prediction model can predict physical property analysis, hazards, and risks of the target substance.

The predictive modeling step (S130) includes Scikit-learn, Keras, Tensorflow, PyTorch, k-Nearest Neighbors (k-NN), decision tree, Random Forest (RF), gradient boosting, ensemble, SVM (Suppoert vector machine), Naive Bayes , Regression, K-means, PCA (principal component analysis), clustering, Grid Search, Cross Validation, ANN, CNN, RNN, Autoencoder, GAN, DNN, QNN, ANOVA (Analysis of variance) Then, the chemical properties having a predetermined similarity range are grouped, and the contribution to the string pattern of the target material is determined from the grouped chemical properties, and classification and regression analysis is performed.

Keras predicts and analyzes the interactions between patterns inside the target substance, Random Forest (RF) calculates the environmental hazards of new substances, and SVM (Suppoert vector machine) analyzes the contributions of hazards and risks according to molecular constituents. , Naive Bayes calculates the probability that the values of conditional environmental hazards and risks do not meet the criteria for chemical evaluation and management according to the molecular structure pattern, and Regression presents the relationship between the amount and toxicity of substances and the maximum amount of non-toxicity, K -means analyzes associations between variables.

Therefore, in the predictive modeling step 130, while predictive analysis using Keras is the main analysis, similar characteristics are clustered through the remaining analysis models, and the influence on the pattern is determined to perform regression analysis. In the case of chemical substances, the properties of the substance can be predicted based on the physical or chemical characteristics such as molecular constituent elements, molecular reactive area, type of bond, steric hindrance, bonding order, and electric density. Because it can be combined, multiple analytical models can be combined to analyze causes affecting toxicity according to the pattern contribution to the target substance.

On the other hand, the predictive modeling step 130 sets the half-lethal concentration (LC50) or the half-number impact concentration (ECx) as a target variable, outputs the smallest residual value through the training data so far, and compares the predicted value with the experimental value. The accuracy of the predicted value can be improved by learning.

The processor determines the toxicity of the target substance by comparing the calculated predicted value with the experimental value of the experiment substance having chemical characteristics similar to the target substance (S140), and matches the determined toxicity with a preset chemical regulation standard, and includes a matching result. Visualize the result of toxicity determination and provide it to the user (S150). The standard for chemical substance regulation may be an item referring to the Act on Registration and Evaluation of Chemical Substances and the Regulation on Specific Methods for Chemical Risk Assessment.

The information providing step (S150) collects the half-lethal concentration (LC50) or the half-water impact concentration (ECx) of the subject, and determines whether the target substance is toxic based on the collected half-lethal concentration or half-water impact concentration.

In addition, the information providing step (S150) analyzes the correlation between the chemical structure and the toxicity of the target material through static analysis, as shown in FIG. 6, for the relationship between the chemical properties or structure of the target material and environmental hazards. The dose-response curve and the predicted value result, and the predetermined chemical regulation criteria (peace law) are matched to provide a probability value for the inability to register the target substance as a licensed substance as a toxicity determination result.

On the other hand, in the information providing step (S150), the chemical structure that reduces the toxicity of the target substance in conjunction with the predicted value may be redesigned to provide the result of the toxicity determination. In the information providing step (S150 ), expert opinions predicting environmental hazards and risks using the physicochemical properties, structures, and chemical structural formulas of the target material may be included in the determination result through the expert terminal.

As described above, the present invention can respond to the possibility of registration of a target substance based on the peace law, and to prepare in advance for an item to be evaluated upon registration, and to perform a redesign of a chemical structural formula that reduces toxicity of a substance in conjunction with a predicted value. have.

Meanwhile, steps S110 to S150 of FIG. 3 may be divided into additional steps or combined into fewer steps according to an embodiment of the present invention.

7 and 8 are diagrams for explaining a prediction model of toxicity based on a convolutional neural network applied to a method for predicting harmfulness and risk of a target substance based on artificial intelligence.

In general, CNN consists of one or several convolutional layers and common artificial neural network layers on top of it, and additionally uses weights and pooling layers. Due to this structure, CNN can fully utilize input data of a two-dimensional structure.

7 and 8, the toxicity prediction model is composed of a plurality of convolutional layers. As the number of convolutional layers increases, more nonlinearity increases, so more useful features can be extracted.

More specifically, it is possible to use cyclic neural network (RNN) techniques such as CNN and LSTM in order to correlate the pattern of chemical structural formulas with the properties of substances. Multiple chemical naming, physical or chemical properties, and biological properties as variables are correlated to extract the characteristics. When the pattern of the chemical formula is analyzed, the extracted pattern characteristic is included in the weight, and the information can be predicted more accurately by inputting information through element-specific operations in the neural network layer.

The dense layer connects both the input and the output, and each connection line includes a weight. Loss is the loss function used to evaluate the current set of weights.

This toxicity prediction model normalizes the data size to roughly match the data size, adds a batch normalization layer before entering a specific hidden layer, corrects the input, and activates the new value as an activation function. Use it in a way that you put it.

As described above, a machine learning or deep learning based toxicity prediction model can predict physical properties, hazards, and risks of a target substance through classification, regression analysis, prediction, and derivation of model results for the target substance. Discrimination results and the like can be visualized.

The embodiments of the present invention described above may also be implemented in the form of a recording medium including instructions executable by a computer, such as program modules, executed by a computer. Such recording media include computer readable media, which can be any available media that can be accessed by a computer and includes both volatile and nonvolatile media, removable and non-removable media. In addition, computer readable media includes computer storage media, which are volatile and nonvolatile implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. , Removable and non-removable media.

The above description of the present invention is for illustration only, and those skilled in the art to which the present invention pertains can understand that the present invention can be easily modified into other specific forms without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as distributed may be implemented in a combined form.

The scope of the present invention is indicated by the following claims rather than the above detailed description, and it should be interpreted that all changes or modified forms derived from the meaning and scope of the claims and equivalent concepts thereof are included in the scope of the present invention. do.

100: physical property prediction device
110: communication module 120: memory
130: processor 140: database

Claims (8)

In the method for predicting the harmfulness and risk of a target material based on artificial intelligence performed by a property prediction device for predicting the property of the target material,
A pattern analysis step of analyzing a character string pattern for a chemical property or structure of the target material;
Provides a prediction model based on artificial intelligence using data on properties and properties of at least one chemical substance as training data, and based on the similarity to the analyzed string pattern and molecular structure characteristics based on the prediction model based on the toxicity prediction model. A predictive modeling step of calculating a predicted value; And
The calculated predicted value is compared with the experimental value of the experimental substance having similar chemical properties to the target substance to determine the toxicity of the target substance, and the determined toxicity is matched with a predetermined chemical substance regulation criterion to obtain physical properties including matching results or A method of predicting the harmfulness and risk of a target substance based on artificial intelligence, comprising the step of providing information providing a result of toxicity determination. According to claim 1,
The predictive modeling step,
Artificial intelligence, in which at least one analysis model is combined to group chemical properties having a predetermined similarity range, and to classify and regression analysis by determining the contribution to the string pattern of the target material from the clustered chemical properties -Based method for predicting the hazard and risk of a target substance. According to claim 2,
The at least one analysis model includes: Scikit-learn, Keras, Tensorflow, PyTorch, k-Nearest Neighbors (k-NN), decision tree, Random Forest (RF), gradient boosting, ensemble, SVM (Suppoert vector machine), Naive Bayes, Regression, K-means, PCA(principal component analysis), clustering, Grid Search, Cross Validation, ANN, CNN, RNN, Autoencoder, GAN, DNN, QNN. Methods for predicting the hazards and risks of a target substance. According to claim 1,
The pattern analysis step,
SMILES (Simplified molecular-input line-entry system) code that converts the chemical structure of the target material into a one-dimensional character array and compares it. A method of predicting the harmfulness and risk of a target substance based on artificial intelligence, which analyzes the similarity between a string pattern and a molecular structure using any one of InCHI and IUPAC. According to claim 1,
The information providing step,
Collect the half-lethal concentration (LC50) or half-water impact concentration (EC50) of the test subject, and numerically predict whether the target substance is toxic, harmful, and risk based on the collected half-lethal concentration or half-water impact concentration and physicochemical properties And, it is to determine whether it meets the chemical regulations, artificial intelligence-based method for predicting the hazard and risk of the target substance. According to claim 1,
The information providing step,
By analyzing the correlation between the chemical structure of the target substance and the hazard and risk through static analysis, the dose-response curve and the predicted value result for the relationship between the chemical structure of the target substance and the environmental hazard and risk, and the preset chemical regulation criteria A method for predicting the harmfulness and risk of a target substance based on artificial intelligence, which provides the legality of a result as a result. According to claim 1,
The preset chemical regulation standard is composed of evaluation criteria for evaluating the target substance as a licensed substance according to the Act on Registration and Evaluation of Chemical Substances, and the method for predicting the hazard and risk of the target substance based on artificial intelligence. According to claim 1,
The information providing step is a method for predicting the harmfulness and risk of a target substance based on artificial intelligence, by redesigning and providing a chemical structure that reduces toxicity of the target substance in conjunction with the predicted value.

Images