A method for predicting the environmental hazards of compounds based on machine learning

Abstract

The present invention discloses a method for predicting the environmental hazard of a compound based on machine learning, wherein the environmental hazard of the compound refers to the persistence, mobility and toxicity of the compound, or the high persistence and high mobility of the compound, and comprises the following steps: (1) establishing a screening standard for the environmental hazard of the compound; (2) collecting sample labels and sample data, and preprocessing the sample data; (3) constructing a prediction model based on a machine learning algorithm, and optimizing the prediction model parameters using the preprocessed sample data; (4) predicting whether the compound to be tested has environmental hazard. The method of the present invention is based on the molecular structure of the compound, and establishes a prediction model according to the relationship between the compound structure and its PMT attribute or vPvM attribute, which fills the gap in the current technology in the field of prediction of PMT substances and vPvM substances, has the advantages of batch, rapid and accurate, and can provide guidance for compound risk assessment and control.

Description

Technical Field

[0001] The present invention relates to the field of compound property prediction, and in particular to a method for predicting the environmental hazard of a compound based on machine learning. Background Art

[0002] With the production and use of new compounds, a large number of chemical substances enter the environmental media every year, with potential ecological and environmental health effects. Among the new compounds put into production and use, there are potential persistent, mobile and toxic (PMT) and very persistent and very mobile (vPvM) properties. How to screen out PMT / vPvM substances is a difficult problem that needs to be solved urgently in the field of environmental risk assessment and control of compounds.

[0003] PMT / vPvM is a concept of material properties. Current research has not disclosed a method for predicting the PMT / vPvM properties of compounds, but the declaration of new compounds and the field of environmental risk assessment and control require the identification of the PMT / vPvM properties of compounds. Therefore, it is urgent to develop a technical method that can predict the PMT / vPvM properties of substances in batches, quickly and accurately, and use this method to screen out PMT / vPvM substances and serve the environmental risk assessment and control of PMT / vPvM substances.

[0004] A Chinese patent document with publication number CN109360610B discloses a method for predicting the biological toxicity of chemical molecules based on a fuzzy neural network, with the hydrophobicity of different chemical structures as the control variable and the biological toxicity as the controlled variable, comprising the following steps: (1) establishing a QSAR model of biological toxicity and octanol / water partition coefficient, that is, establishing a model of biological toxicity and octanol / water partition coefficient for the synthesis process of chemical molecules; (2) establishing a fuzzy neural network chemical molecule toxicity prediction model; (3) establishing a genetic algorithm model for optimizing NFN parameters for correcting fuzzy neural network parameters; (4) finally using the optimized fuzzy neural network model to calculate and predict the biological toxicity value of the new molecule.

[0005] A Chinese patent document with publication number CN112541088A discloses a method for constructing a hazardous compound library based on a knowledge graph, including: (1) constructing a hazardous compound data set; (2) determining the basic properties of the compound based on the content of the compound knowledge graph, and determining the accident risk information to be used; (3) establishing a risk information extraction model; (4) annotating the data and using a neural network model with AlBert as a pre-trained model to extract information; (5) integrating the data and constructing a knowledge graph. The invention stores various types of information about compounds in the knowledge base in the form of a knowledge graph, which is more convenient for reference. Summary of the invention

[0006] The present invention provides a method for predicting the environmental hazards of compounds based on machine learning, which fills the gap in the current technology in the field of prediction of PMT substances and vPvM substances. It has the advantages of batch, fast and accurate, and can provide guidance for compound risk assessment and control.

[0007] The specific technical solutions adopted are as follows:

[0008] A method for predicting the environmental hazard of a compound based on machine learning, wherein the environmental hazard of the compound refers to the persistence, mobility and toxicity of the compound, or the high persistence and high mobility of the compound, specifically comprising the following steps:

[0009] (1) Establishing environmental hazard screening criteria for compounds;

[0010] (2) extracting some compounds from the compound database as samples, screening the samples according to the screening criteria established in step (1), and using the screening results as sample labels; exporting the smiles expression of the samples, extracting the specific molecular features of the samples as sample data, and preprocessing the sample data;

[0011] (3) Build a prediction model based on a machine learning algorithm, and use the preprocessed sample data to train the prediction model under the supervision of sample labels to optimize the prediction model parameters;

[0012] (4) After deriving the smiles expression of the test compound, the specific molecular characteristics of the test compound are extracted and used as the test data to input into the parameter optimization prediction model to predict whether the test compound has environmental hazards.

[0013] The method of the present invention is based on the molecular structure of the compound, and establishes a prediction model according to the relationship between the compound structure and its PMT properties (persistence, mobility and toxicity) or vPvM properties (high persistence and high mobility). No animal or cell experiments are required, and the PMT property or vPvM property of the compound can be predicted by the established prediction model based only on the compound structure. The method is rapid, economical, and can be processed in large quantities, providing a basis for the evaluation and management of PMT / vPvM substances.

[0014] There is no clear PMT / vPvM substance category in the existing database. The present invention establishes PMT / vPvM substance screening standards and checks and screens compounds based on the physical and chemical properties of the compounds.

[0015] The persistence screening criteria for compounds include: compounds with a degradation half-life of greater than 60 days in seawater at a pH of 6-8 and a temperature of 9°C; or compounds with a degradation half-life of greater than 40 days in freshwater or estuarine waters at a pH of 6-8 and a temperature of 12°C; or compounds with a degradation half-life of greater than 180 days in marine sediments at a pH of 6-8 and a temperature of 9°C; or compounds with a degradation half-life of greater than 120 days in freshwater sediments or estuarine sediments at a pH of 6-8 and a temperature of 12°C; or compounds with a degradation half-life of greater than 120 days in soil at a pH of 6-8 and a temperature of 12°C.

[0016] The compound mobility screening criteria include: compounds with a solubility of ≥150 μg / L in water with a pH of 6 to 8 and a temperature of 12°C and a logarithm of the n-octanol-water partition coefficient (log Koc) ≤4.5.

[0017] The toxicity screening criteria for compounds include: compounds with no long-term detectable effect concentrations <0.01 mg / L in marine or freshwater organisms; or compounds that have been proven experimentally to have at least one of the following properties: carcinogenicity, germ cell mutagenicity, reproductive toxicity, and chronic toxicity; or compounds with a Derived No-Effect Level (DNEL) ≤9 μg / kg / d.

[0018] The high persistence screening criteria for compounds include: substances with a degradation half-life of more than 60 days in freshwater or estuarine waters at a pH of 6-8 and a temperature of 12°C; or substances with a degradation half-life of more than 180 days in freshwater sediments or estuarine sediments at a pH of 6-8 and a temperature of 12°C; or substances with a degradation half-life of more than 180 days in soils at a pH of 6-8 and a temperature of 12°C.

[0019] The high mobility screening criteria for compounds include: compounds with a solubility of ≥150 μg / L in water with a pH of 6 to 8 and a temperature of 12°C and a logarithm of the n-octanol-water partition coefficient (log Koc) ≤3.0.

[0020] Preferably, the compound databases include PubChem, ECHA, ChemDB and ChEMBL.

[0021] Preferably, in step (2), the compounds are screened according to the screening criteria established in step (1). If the compounds meet the environmental hazard screening criteria of the compounds, that is, the compounds simultaneously meet the persistence screening criteria, the mobility screening criteria, and the toxicity screening criteria, or the compounds simultaneously meet the high persistence screening criteria and the high mobility screening criteria, then the sample label of the compound is 1 (positive); otherwise, the sample label is 0 (negative).

[0022] The specific molecular features include molecular descriptors and / or one or more molecular fingerprints, which can be extracted using software such as Dragon, ChemDes, and RDKit.

[0023] The preprocessing method includes standardization and normalization. Standardization is to process data according to the columns of the feature matrix and convert the eigenvalues ​​of the samples to the same dimension; normalization is to process data according to the rows of the feature matrix and map the data to a specified range.

[0024] In step (3), the machine learning algorithm is selected from a random forest algorithm, a logistic regression algorithm, a naive Bayesian algorithm, a support vector machine algorithm, and a deep learning algorithm.

[0025] Further preferably, in step (3), the machine learning algorithm is a support vector machine algorithm, and a Gaussian radial basis kernel function is used as a support vector machine parameter to construct a prediction model. The Gaussian radial basis kernel function formula is:

[0026]

[0027] Where, parameter γ = –1 / 2σ 2 , where x is a point in space, x' is the center of the kernel function, ||x-x'|| represents the Euclidean distance from the point to the center of the kernel function, σ is the width parameter of the function, and K(x, x') is a monotonic function of the Euclidean distance between any point x in space and the center x' of the kernel function.

[0028] Preferably, in step (3), the sample data is divided into a training set and a test set, the ratio of the training set to the test set is preferably 7:3, the training set is used to train the prediction model, and the test set is used to evaluate the goodness of the prediction model. During the training process, the parameter γ of the Gaussian radial basis kernel function is set at 10 -5 50 values ​​are taken at equal intervals in the range from 1 to 1 to optimize the prediction model parameters;

[0029] The model goodness evaluation indicators include: accuracy, recall rate and area under the curve AUC value. Accuracy refers to the proportion of all correctly predicted sample data to the total sample data; recall rate refers to the proportion of correctly predicted positive samples to all actual positive samples; AUC value refers to the area under the ROC curve, which is a curve with false positive rate as the horizontal axis and true positive rate as the vertical axis.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] (1) The present invention fills the gap in the current technology in the field of PMT substance and vPvM substance prediction by establishing screening criteria and using sample data and sample labels to construct an optimized prediction model based on a machine learning algorithm, providing a batch, fast, and accurate PMT substance and vPvM substance prediction method with an accuracy of 84.5%.

[0032] (2) The present invention has a wide range of applications and can predict the environmental hazards of compounds in batches, quickly and accurately. Based on the Smiles expression of the substance to be tested, it can predict in batches, quickly and accurately whether the compound is a PMT substance or a vPvM substance, providing guidance for compound risk assessment and control. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of the method for predicting the environmental hazard of a compound based on machine learning in the present invention.

[0034] Figure 2 It is a functional relationship diagram of the parameter γ of the Gaussian radial basis kernel function and the accuracy of the prediction model. DETAILED DESCRIPTION

[0035] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0036] The flowchart of the method for predicting the environmental hazard of a compound based on machine learning in the embodiment of the present invention is as follows Figure 1 As shown, it includes 4 steps as shown in the figure:

[0037] (1) Establishing a screening standard for the environmental hazard of a compound. In the present invention, environmental hazard refers to the persistence, mobility and toxicity of a compound, or the high persistence and high mobility of a compound.

[0038] The persistence screening criteria for compounds include: compounds with a degradation half-life of more than 60 days in seawater at a pH of 6-8 and a temperature of 9°C; or compounds with a degradation half-life of more than 40 days in freshwater or estuarine waters at a pH of 6-8 and a temperature of 12°C; or compounds with a degradation half-life of more than 180 days in marine sediments at a pH of 6-8 and a temperature of 9°C; or compounds with a degradation half-life of more than 120 days in freshwater sediments and estuarine sediments at a pH of 6-8 and a temperature of 12°C; or compounds with a degradation half-life of more than 120 days in soil at a pH of 6-8 and a temperature of 12°C.

[0039] The screening criteria for compound mobility include: compounds with a solubility of ≥150 μg / L and a logarithm of the n-octanol-water partition coefficient (log Koc) ≤4.5 in water with a pH of 6-8 and a temperature of 12°C.

[0040] The screening criteria for compound toxicity include: compounds with no long-term detected effect concentration <0.01 mg / L in marine or freshwater organisms; or compounds that have been proven to have at least one of the following properties: carcinogenicity, germ cell mutagenicity, reproductive toxicity, and chronic toxicity through experiments; or compounds with a Derived No-Effect Level (DNEL) ≤9 μg / kg / d.

[0041] The high persistence screening criteria for compounds include: substances with a degradation half-life of more than 60 days in freshwater or estuarine waters at a pH of 6-8 and a temperature of 12°C; or substances with a degradation half-life of more than 180 days in freshwater sediments or estuarine sediments at a pH of 6-8 and a temperature of 12°C; or substances with a degradation half-life of more than 180 days in soils at a pH of 6-8 and a temperature of 12°C.

[0042] The high mobility screening criteria for compounds include: compounds with a solubility of ≥150 μg / L in water with a pH of 6 to 8 and a temperature of 12°C and a logarithm of the n-octanol-water partition coefficient (log Koc) ≤3.0.

[0043] (2) extracting some compounds as samples, screening the samples according to the screening criteria established in step (1), and using the screening results as sample labels; exporting the smiles expression of the samples, extracting the specific molecular features of the samples as sample data, and preprocessing the sample data; the preprocessing includes standardization and normalization.

[0044] 120 compounds were extracted from the PubChem open source database as samples, including POPs (Persistent Organic Pollutants), suspected POPs and other organic substances.

[0045] The samples were screened according to the screening criteria established in step (1), and the screening results were used as sample labels; if the sample met the screening criteria for the environmental hazard of the compound, the sample label was 1, otherwise the sample label was 0. The Smiles expression and linear molecular structure of the sample were exported from the PubChem database, and the specific structural features of the sample were extracted using the RDKit software, with the linear molecular structure of the sample as input and the RDKit molecular descriptor as output. Each column of data corresponded to a molecular descriptor, and finally 208 columns of molecular descriptors were obtained, i.e., a feature matrix of 120 rows and 208 columns. The number 1 indicated that it was a PMT / vPvM substance, and the number 0 indicated that it was a non-PMT / vPvM substance. A sample data set was established, and the sample data set was standardized and normalized.

[0046] (3) Construct a prediction model based on a machine learning algorithm, and use the preprocessed sample data to train the prediction model under the supervision of sample labels to optimize the prediction model parameters.

[0047] The sample data is randomly divided into a test set and a training set, where the ratio of the training set to the test set is 7:3. The machine learning algorithm uses the binary classifier (SVC) in the support vector machine algorithm and uses the Gaussian radial basis kernel function.

[0048] The Gaussian radial basis kernel function formula is:

[0049]

[0050] Where, parameter γ = –1 / 2σ 2 , where x is a point in space, x' is the center of the kernel function, ||x-x'|| represents the Euclidean distance from the point to the center of the kernel function, σ is the width parameter of the function, and K(x, x') is a monotonic function of the Euclidean distance between any point x in space and the center x' of the kernel function.

[0051] Use the training set to train the prediction model, use the test set to evaluate the prediction model, and optimize the prediction model parameters; use the loop statement to adjust the parameter γ between 10 and 10. -5 50 values ​​are taken at equal intervals in the range of 1 to 1. At the same time, the accuracy of the model is output in each cycle, and the accuracy and the corresponding γ value are listed. After the cycle is completed, the maximum accuracy and the corresponding γ value in the list are output, and the accuracy of the prediction model is used as the vertical coordinate and the parameter γ is used as the horizontal coordinate to output the corresponding function curve graph of the two, as shown in the figure below: Figure 2 shown.

[0052] The support vector machine algorithm can improve the generalization ability of the learning machine by seeking to minimize the structured risk, and minimize the empirical risk and the confidence range, so as to achieve the purpose of obtaining good statistical laws even when the statistical sample size is small. There is no clear PMT / vPvM substance category in the existing database. The present invention establishes screening criteria and checks and screens compounds according to the physical and chemical properties of the compounds. Even when the sample size is small, good statistical results can still be obtained, the prediction accuracy is high, and the cost and operation process of the method are saved.

[0053] (4) After deriving the smiles expression of the test compound, the specific molecular characteristics of the test compound are extracted and used as the test data to input into the parameter optimization prediction model to predict whether the test compound has environmental hazards.

[0054] 32 organic compounds (mainly including fluorine-containing and chlorine-containing organic compounds) were selected to derive their Smiles expressions, and batch prediction was performed using the method of the present invention. The prediction results are shown in Table 1:

[0055] Table 1 Prediction results of PMT properties / vPvM properties of 32 organic compounds

[0056]

[0057] In Table 1, 1 indicates positive, 0 indicates negative, and the accuracy of the method reaches 84.4%. That is, the method of the present invention has a high accuracy in actually predicting the PMT property or vPvM property of the compound, which can provide an important basis for compound risk assessment and control.

[0058] The embodiments described above provide a detailed description of the technical solutions of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements or similar substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting the environmental hazard of a compound based on machine learning, characterized in that: The environmental hazard of the compound refers to the persistence, mobility and toxicity of the compound, or the high persistence and high mobility of the compound, which specifically includes the following steps: (1) Establishing screening criteria for the environmental hazards of compounds; (2) extracting some compounds from the compound database as samples, screening the samples according to the screening criteria established in step (1), and using the screening results as sample labels. If the samples meet the screening criteria for the environmental hazards of the compounds, the sample label is 1, otherwise the sample label is 0; after exporting the smiles expression of the samples, the specific molecular features of the samples are extracted as sample data, and the sample data is preprocessed; the specific molecular features include molecular descriptors and / or one or more molecular fingerprints; (3) Build a prediction model based on a machine learning algorithm, and use the preprocessed sample data to train the prediction model under the supervision of sample labels to optimize the prediction model parameters; (4) After deriving the smiles expression of the test compound, the specific molecular features of the test compound are extracted and used as the test data to input into the parameter optimization prediction model to predict whether the test compound has environmental hazards; A compound meeting the environmental hazard screening criteria of a compound means that the compound simultaneously meets the persistence screening criteria, the mobility screening criteria and the toxicity screening criteria, or the compound simultaneously meets the high persistence screening criteria and the high mobility screening criteria.

2. The method for predicting the environmental hazard of a compound based on machine learning according to claim 1, characterized in that: The compound databases include PubChem, ECHA, ChemDB and ChEMBL.

3. The method for predicting the environmental hazards of compounds based on machine learning according to claim 1, characterized in that: In step (3), the machine learning algorithm is selected from a random forest algorithm, a logistic regression algorithm, a naive Bayes algorithm, a support vector machine algorithm or a deep learning algorithm.

4. The method for predicting the environmental hazards of compounds based on machine learning according to claim 3, characterized in that: In step (3), the machine learning algorithm is a support vector machine algorithm, and the Gaussian radial basis kernel function is used as a support vector machine parameter to construct a prediction model.

5. The method for predicting the environmental hazards of compounds based on machine learning according to claim 1, characterized in that: In step (3), the sample data is divided into a training set and a test set, the prediction model is trained using the training set, and the prediction model is evaluated using the test set to optimize the prediction model parameters.

6. The method for predicting the environmental hazards of compounds based on machine learning according to claim 5, characterized in that: During the training process, the parameter γ of the Gaussian radial basis kernel function is set at 10 -5 The prediction model parameters are optimized by taking values ​​with equal intervals in the range from 1 to 1.

Images