Aquatic organism toxic effect threshold value calculation method based on machine learning
By constructing a database of chemical target toxicity endpoints and toxicity action modes through machine learning, the threshold of toxic effects on aquatic organisms is predicted. This solves the problem of calculating the threshold of mixed toxic effects of chemical pollutants in complex aquatic environments, and enables rapid and accurate calculation of toxicity thresholds and water quality risk management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV
- Filing Date
- 2025-12-08
- Publication Date
- 2026-04-21
AI Technical Summary
Existing technologies struggle to quickly and accurately calculate the combined toxicity thresholds of chemical pollutants in complex aquatic environments. Traditional methods are time-consuming, labor-intensive, and lack sufficient data. Furthermore, traditional targeted chemical analysis methods cannot explain the combined toxicity effects of low-concentration pollutants.
Machine learning methods were used to construct a database of chemical target toxicity endpoint features and a database of toxic action modes. Through feature screening and model training, the target biological toxicity test endpoints and toxic action modes were predicted, reference compounds were screened, and the toxicity effect thresholds of aquatic organisms were calculated.
It enables rapid and accurate calculation of toxicity thresholds for aquatic organisms, saving time and costs, enriching the acute toxicity database of biological tests in aquatic organisms, and providing a new method for water quality risk management.
Smart Images
Figure CN121905346A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of compound environmental risk assessment methods, specifically to a machine learning-based method for calculating the threshold of toxic effects on aquatic organisms. Background Technology
[0002] Rapid industrialization and urbanization have led to a large influx of chemical pollutants into natural water bodies. These pollutants are diverse in type and structure, and exhibit various biotoxic hazards, posing a severe challenge to risk assessment and control. While traditional targeted chemical analysis methods can quantitatively determine the concentration of specific pollutants, they often overlook the potential combined toxic effects of pollutants below the detection limit or at safe concentrations when mixed. Studies have reported that detecting the concentrations of 290 common, known hazardous chemicals in streams could only explain <8% of the contribution to biotoxic effects.
[0003] Effect-based methods (EBMs) can directly detect the combined toxic effects of mixtures in environmental samples, providing more comprehensive and accurate toxicity information. However, in practical applications, these methods generally lack effect thresholds for determining the acceptable toxicity of water. To address this, researchers have proposed effect-based trigger values (EBTs) to determine the potential health risks posed by water quality. When the toxicity value tested using a particular EBM method exceeds the corresponding threshold, it indicates a potential hazard to human health or ecological safety, necessitating a detailed and in-depth chemical screening and identification process.
[0004] Effect trigger values (EBTs) are specific to both bioassay methods and toxicity endpoints; that is, an EBT targets only a specific toxicity endpoint. Typically, the occurrence of a certain biotoxicity endpoint induced by a water sample is the result of the combined action of a group of chemicals with the same mode of toxicity. Therefore, the core of the effect trigger value theory is to identify a strongly positive chemical for a specific toxicity endpoint of a particular bioassay method, using it as a reference compound. The toxicity efficacy of other positive chemicals is then converted into the equivalent concentration of the reference compound, and the EBT is derived through additive effects or other calculation formulas. The key to this process is accurately determining the toxic mode of action of the compound and selecting a reference compound with the same toxic mode of action as the sample.
[0005] However, pollutants in real-world aquatic environments are diverse and their toxic action modes vary. For specific toxic endpoints (such as aromatic hydrocarbon receptor activation), the toxic action mode is singular, and the determination of the reference compound is relatively simple; but in most cases, the toxic action mode of a compound is unclear. Traditional methods relying on animal experiments to determine the toxic action mode are time-consuming and labor-intensive. Computational toxicology methods, which predict toxicological effects and action modes based on the molecular structure of compounds, can effectively save time and costs. However, challenges remain in model construction, including a lack of toxicity data, inconsistent testing conditions, and diverse endpoints. Summary of the Invention
[0006] Purpose of the invention: The technical problem to be solved by the present invention is to provide a method for calculating the threshold of toxic effects of aquatic organisms based on machine learning, which addresses the shortcomings of the existing technology.
[0007] To address the aforementioned technical problems, this invention discloses a machine learning-based method for calculating the threshold of toxic effects on aquatic organisms, comprising:
[0008] Step 1: Construct a database of chemical target toxicity endpoint features, which includes chemical molecule descriptor information and target biotoxicity test results information, and preprocess the feature data;
[0009] Step 2: Establish and optimize the prediction model for the target toxicity endpoint of chemicals. Based on the database established in Step 1, use the feature screening method to select chemical molecular descriptors as input; construct a regression model, use the screened molecular descriptors as input, predict the endpoint of the target biotoxicity test, and construct the prediction model for the endpoint of the target toxicity test.
[0010] Step 3: Construct a chemical toxicity mode database. The database includes chemical molecule descriptor information and corresponding toxicity mode characteristics. The data is then preprocessed.
[0011] Step 4: Establish a chemical toxicity mode prediction model. Based on the database established in Step 3, use principal component analysis to screen chemical molecule descriptors as input; input the screened molecule descriptors into the classification model to predict their toxicity mode and construct a toxicity mode classification model.
[0012] Step 5: Based on the target toxicity endpoint prediction model and toxicity mode prediction model obtained in Step 2 and Step 4, supplement the compound toxicity characteristic parameters and screen reference compounds of the target toxicity test type in chemical mixtures.
[0013] Step 6: Based on the reference compounds screened out in Step 5, calculate the toxicity effect threshold for the target toxicity test of aquatic organisms, and complete the rapid calculation of the chemical toxicity effect threshold based on machine learning, which is used for chemical aquatic risk identification and control.
[0014] Furthermore, the construction of the chemical target toxicity endpoint characterization database described in step 1 includes:
[0015] Step 1-1: Collect and organize data on the endpoints of chemical toxicity tests for aquatic organisms from literature and databases;
[0016] Steps 1-2: Based on the SMILES structure of the chemical, calculate the corresponding 2D molecular descriptor, remove missing values, outliers and invalid values, eliminate descriptor columns with variance of 0, and perform standardization.
[0017] Furthermore, the establishment and optimization of the chemical target toxicity endpoint prediction model described in step 2 includes:
[0018] Step 2-1: Calculate the distance between molecular descriptor vectors, cluster molecular descriptors with correlation ≥ 0.95, retain only the molecular descriptor with the largest variance in each cluster, and use a filtering method to select molecular descriptors as input.
[0019] Step 2-2: Divide the dataset into a test set and a validation set according to the proportion, and perform internal validation using five-fold cross-validation. Take the molecular descriptors screened out in Step 2-1 as input and the endpoint data of the toxicity test as output. Use regression algorithms including random forest, extreme gradient boosting tree, support vector regression, K-nearest neighbor algorithm, voting ensemble learning, stacking ensemble learning, Bagging ensemble learning, AdaBoost ensemble learning and multiple linear regression to train and build the model.
[0020] Steps 2-3: Use mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). 2 Mean absolute percentage error (MAPE) characterizes the model's predictive performance.
[0021] Steps 2-4: Draw Williams diagrams based on standardized residuals and lever distances to represent the application domain;
[0022] Steps 2-5, with 𝑅 2 >0.500 and MAPE<100% are the criteria for judging the model performance as qualified. The best predictive model for the target biotoxicity test endpoint is determined based on the model performance.
[0023] Furthermore, the construction of the chemical toxicity mode database described in step 3 includes:
[0024] Step 3-1: Collect and organize the CAS, CID, SID, and SMILES structural information of organic chemicals from literature and databases;
[0025] Step 3-2: Based on the SMILES structure of chemicals, chemicals are classified into four categories: inert chemicals (Class 1), weakly inert chemicals (Class 2), reactive chemicals (Class 3), and specific chemicals (Class 4).
[0026] Step 3-3: Remove missing values, outliers, and invalid values, eliminate descriptor columns with a variance of 0, and perform standardization.
[0027] Furthermore, the establishment of a chemical toxicity pattern prediction model described in step 4 includes:
[0028] Step 4-1: Perform principal component analysis on the input variables and obtain the contribution of each principal component. Select the top 50 principal components as input features.
[0029] Step 4-2: Divide the dataset into a test set and a validation set according to the proportion. Take the input features screened out in Step 4-1 as input and the chemical toxicity mode information as output. Use classification models including decision tree, random forest, support vector machine, Gaussian Naive Bayes model, logistic regression model and voting method for training.
[0030] Step 4-3: Use the area under the ROC curve (AUC) and confusion matrix, and select accuracy, precision, recall, and F1 score to measure the results.
[0031] Step 4-4: Determine the best classification scheme based on the performance of the classification model. Analyze the performance of the four-category toxicity pattern classification model based on the evaluation indicators of the classification model, including inert chemicals, weakly inert chemicals, reactive chemicals, and chemicals with specific toxicity patterns. Optimize the prediction model based on the classification performance, and continue to use the adjusted scheme for model training to evaluate the model performance.
[0032] Furthermore, the optimization of the prediction model based on the classification effect described in step 4-4 includes: analyzing the accuracy of the constructed classification model, merging inert chemicals (Class 1) and weakly inert chemicals (Class 2) into non-specific chemicals, and classifying chemicals into three categories in total: non-specific chemicals, reactive chemicals, and specific chemicals.
[0033] Furthermore, the reference compound for screening the target toxicity test type of the chemical mixture described in step 5 includes:
[0034] Step 5-1: Collect the CAS of the substances contained in the chemical mixture and calculate the SMILES structural formula of the chemical.
[0035] Step 5-2: Based on the chemical target toxicity endpoint prediction model obtained in Step 2, expand the data for the biotoxicity test endpoints of substances in chemical mixtures.
[0036] Step 5-3: Based on the chemical toxicity mode prediction model obtained in Step 4, the toxicity mode information of substances in the chemical mixture is expanded.
[0037] Step 5-4: For the target biotoxicity test, prioritize the chemical with the lowest toxicity endpoint value among the specific category chemicals as the reference compound; if the toxicity endpoint value of the specific toxic chemical is unavailable, select the chemical with the lowest toxicity endpoint value among the reactive category chemicals as the reference compound; if only the toxicity endpoint data of the non-specific category chemicals is available, directly select the chemical with the lowest toxicity endpoint value among them as the reference compound.
[0038] Furthermore, the calculation of the toxicity effect threshold for the aquatic organism target toxicity test in step 6 includes:
[0039] Step 6-1: Using the EBTs calculation method proposed by EBT-EQS, calculate the effect-based trigger values (EBTs) for the target biotoxicity test type of the chemical mixture.
[0040] Step 6-2: Use whether the difference between the predicted value and the actual value of EBT-BEQ / n in the same type of biotoxicity test exceeds one order of magnitude as the standard for evaluating the predicted value of EBTs, and evaluate the accuracy of the calculation method.
[0041] Step 6-3: If the difference between the predicted value and the true value of EBT-BEQ / n exceeds one order of magnitude, calculate EBT-logBEQ using logBEQ, perform statistical analysis on EBT-BEQ and EBT-logBEQ, calculate EBT-BEQ using logBEQ based on the linear relationship between the two, and evaluate the accuracy of the calculation results according to the rules in Step 6-2.
[0042] Furthermore, the EBT-BEQ / n mentioned in step 6-2 refers to the EBT calculated based on the following formula:
[0043]
[0044] in, Let represent the bioeffect equivalent of the i-th chemical, and n represent the quantity of the chemical.
[0045] Further, step 6-3, which calculates EBT-BEQ using logBEQ based on the linear relationship between EBT-BEQ and EBT-logBEQ, includes:
[0046] Linear fitting was performed on the true values of the corresponding biotoxicity tests EBT-BEQ and EBT-logBEQ. The significance of the regression analysis results was tested by regression error analysis. Based on the linear analysis results, the mathematical relationship between EBT-BEQ and logBEQ was obtained, and EBT-BEQ based on logBEQ was calculated.
[0047] Beneficial effects:
[0048] This invention provides a machine learning-based method for calculating the threshold of toxic effects on aquatic organisms. By using machine learning, it overcomes the limitations of incomplete public data on the toxic effects of large quantities of chemicals and unknown toxic action modes, enabling the scientific calculation of the toxicity threshold of target organisms. This expands a new approach to the derivation of EBTs, saves time and economic costs, and provides a method for the assessment and control of aquatic risks of chemicals. Attached Figure Description
[0049] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.
[0050] Figure 1 This is a schematic diagram of the method flow of the present invention.
[0051] Figure 2 This diagram illustrates the accuracy results of the four-classification and three-classification methods for the toxicity modes of compounds.
[0052] Figure 3 This is a schematic diagram showing the results of the reference compound selection. Detailed Implementation
[0053] The embodiments of the present invention will now be described with reference to the accompanying drawings.
[0054] Etiologically sensitive biological tests (EBTs) are crucial for managing the ecological risks of mixed pollutant systems, but their practical application is severely hampered by insufficient data on pollutant toxicity mechanisms and chemical properties. In recent years, machine learning, with its efficient data mining and pattern recognition capabilities, has provided new tools for predicting compound toxicity and elucidating mechanisms. This method not only significantly reduces the time and cost required by traditional experimental methods but also predicts compound toxicity characteristics and toxic action patterns even when experimental data is scarce. Therefore, this invention proposes a machine learning-based EBTs derivation and calculation method, focusing on both multi-species acute toxicity numerical prediction and toxic action pattern prediction. This aims to provide data support for the derivation of EBTs in in vivo biological tests, enriching the acute toxicity database for in vivo biological tests in aquatic organisms. Furthermore, it optimizes the threshold derivation calculation formula and scheme, thus providing a new method for water quality risk management based on EBTs thresholds.
[0055] This application provides a machine learning-based method for calculating the threshold of toxic effects on aquatic organisms, including the following steps:
[0056] Step 1: Construct a database of chemical target toxicity endpoint characteristics. These characteristics include chemical molecular descriptor information and target biotoxicity test results. The data is then preprocessed. Specifically, this includes:
[0057] Step 1-1: Collect and organize data on the endpoints of chemical toxicity tests on typical aquatic organisms from literature and databases (CHRIP, PubChem, etc.);
[0058] Steps 1-2: Based on the SMILES structure of the chemical, use Python Mordred to calculate the corresponding 2D molecular descriptors, remove missing values, outliers and invalid values, eliminate descriptor columns with variance of 0, and perform standardization.
[0059] Step 2: Establish and optimize the target toxicity endpoint prediction model for chemicals. Based on the database established in Step 1, use feature screening to select chemical molecular descriptors as input; construct a regression model, using the screened molecular descriptors as input, to predict the results of the target biotoxicity test endpoint, thus constructing the target toxicity endpoint prediction model; specifically including:
[0060] Step 2-1: Calculate the distance between molecular descriptor vectors, cluster molecular descriptors with correlation ≥ 0.95, retain only the molecular descriptor with the largest variance in each cluster, and use a filtering method to select molecular descriptors as input.
[0061] Step 2-2: Divide the dataset into a test set and a validation set in a 4:1 ratio, and perform internal validation using five-fold cross-validation. Use the molecular descriptors selected in Step 2-1 as input and the endpoint data of the toxicity test as output. Train and build the model using regression algorithms (random forest, extreme gradient boosting tree, support vector regression, K-nearest neighbor algorithm, voting ensemble learning, stacking ensemble learning, bagging ensemble learning, AdaBoost ensemble learning, and multiple linear regression).
[0062] Steps 2-3 use mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The mean absolute percentage error (MAPE) characterizes the model's predictive performance, as detailed below:
[0063] (1)
[0064] (2)
[0065] (3)
[0066] (4)
[0067] Where n represents the quantity of chemicals, y i It is the endpoint value of chemical toxicity test obtained from the experiment. i is the endpoint value of chemical toxicity test obtained from the predictive model, and ȳ is the average value of the endpoint value of chemical toxicity test obtained from the experiment;
[0068] Steps 2-4, based on standardized residuals and lever distance (ℎ 𝑖 Draw a Williams diagram to represent the application domain. 𝑖 The calculation method is as follows:
[0069] (5)
[0070] (6)
[0071] Where X is a matrix of multiple chemicals and multiple features in the training set, x i This represents the molecular descriptor vector for the i-th chemical, ℎ 𝑖 Let ℎ be the lever distance for the i-th chemical. ∗ The warning value for the defined lever distance, where ℎ is the number of molecular descriptors, n is the number of chemicals, and ℎ 𝑖 > ℎ ∗ This indicates that the structure of this chemical is significantly different from that of other chemicals.
[0072] The aforementioned standard residuals specifically include:
[0073] (7)
[0074] Steps 2-5, with 𝑅 2 >0.500 and MAPE<100% are the criteria for judging the model performance as qualified. The best predictive model for the target biotoxicity test endpoint is determined based on the model performance.
[0075] Step 3: Construct a database of chemical toxicity modes, including chemical molecular descriptor information and corresponding toxicity mode characteristics, and preprocess the data; specifically including:
[0076] Step 3-1: Collect and organize the CAS, CID, SID, and SMILES structural information of organic chemicals from literature and databases (TOX21, PubChem, EBT dataset, etc.);
[0077] Step 3-2: Based on the SMILES structural formula of the chemical, classify the chemical into four categories according to the Verhaar classification method: Class 1 (inert), Class 2 (weakly inert), Class 3 (reactive), and Class 4 (specific).
[0078] Step 3-3: Remove missing values, outliers, and invalid values, eliminate descriptor columns with a variance of 0, and perform standardization.
[0079] Step 4: Establish a chemical toxicity pattern recognition model. Based on the database established in Step 3, principal component analysis is used to screen chemical molecule descriptors as input. The screened molecule descriptors are then input into a classification model to predict their toxicity patterns, thus constructing a toxicity pattern classification model. Specifically, this includes:
[0080] Step 4-1: Perform principal component analysis on the input variables and obtain the contribution of each principal component. Select the top 50 principal components as input features.
[0081] Step 4-2: Divide the dataset into a test set and a validation set in a 4:1 ratio. Use the input features selected in Step 4-1 as input and the chemical toxicity mode information as output. Train the dataset using classification algorithms (decision tree, random forest, support vector machine, Gaussian Naive Bayes model, logistic regression model, and voting method).
[0082] Step 4-3 uses the area under the ROC curve (AUC) and confusion matrix, and selects accuracy, precision, recall, and F1 score to measure the results, as follows:
[0083] (8)
[0084] (9)
[0085] (10)
[0086] (11)
[0087] Among them, TP is a true positive, that is, a positive chemical that is accurately classified; TN is a true negative, that is, a negative chemical that is correctly classified; FP is a false positive, that is, a positive chemical that is misclassified; and FN is a false negative, that is, a negative chemical that is misclassified.
[0088] Step 4-4: Determine the optimal classification scheme based on the performance of the classification model. Using the model's performance evaluation indicators as a basis, analyze the effectiveness of the four-category toxicity action pattern classification model (inert chemicals, weakly inert chemicals, reactive chemicals, and chemicals with specific toxicity action patterns). Adjust the scheme based on the classification results. Use the adjusted scheme for further model training and evaluate the model's performance.
[0089] The adjustment of the scheme based on the classification effect described in step 4-4 specifically includes: analyzing the accuracy of the constructed classification model, merging Class 1 (inert chemicals) and Class 2 (weakly inert chemicals) into a non-specific chemical category, and classifying chemicals into three categories in total: non-specific chemicals, reactive chemicals, and specific chemicals.
[0090] Step 5: Based on the target toxicity endpoint prediction model and toxicity pattern recognition model obtained in Steps 2 and 4, supplement the compound toxicity characteristic parameters and screen reference compounds for the target toxicity test type in chemical mixtures; specifically including:
[0091] Step 5-1: Collect the CAS of the substances contained in the chemical mixture and calculate the SMILES structure of the chemicals based on PubChem;
[0092] Step 5-2: Based on the chemical target toxicity endpoint prediction model obtained in Step 2, expand the data for the biotoxicity test endpoints of substances in chemical mixtures.
[0093] Step 5-3: Based on the chemical toxicity mode prediction model obtained in Step 4, the toxicity mode information of substances in the chemical mixture is expanded.
[0094] Step 5-4: For the target biotoxicity test, prioritize the chemical with the highest toxicity (lowest toxicity endpoint value) among the specific toxicity categories as the reference compound; if the toxicity endpoint value of the specific toxicity chemical is unavailable, select the chemical with the highest toxicity among the reactive toxicity categories as the reference compound; if only the toxicity endpoint data of the non-specific toxicity categories are available, directly select the chemical with the highest toxicity among them as the reference compound.
[0095] Step 6: Based on the reference compounds screened out in Step 5, calculate the toxicity effect threshold of the target toxicity endpoint for aquatic organisms, completing the rapid calculation of the chemical toxicity effect threshold based on machine learning, which is used for chemical aquatic risk identification and control; specifically including:
[0096] Step 6-1: Using the EBTs calculation method proposed by EBT-EQS, calculate the effect-based trigger values (EBTs) for the target biotoxicity test type of the chemical mixture; the specific EBTs calculation method is as follows:
[0097] (12)
[0098] (13)
[0099] In order to convert chemical concentration into corresponding biological effect equivalent (BEQ), the chemical substance i and its relative potency (REP) to the reference compound are first obtained, where REP... i The calculation formula is formula (12). This indicates the toxicity effect value (EC50) of the reference compound. 50 LC 50 wait), The toxicity effect value of compound i is expressed; then, the concentration of chemical substance i in the bioassay is expressed using formula (13). Multiply by REP i Get BEQ i .
[0100] EBT(M) and EBT(μg / L) under different conditions were calculated using the following methods, where Option E was used only for calculating EBT(M), and the other four methods were used for calculating both EBT(M) and EBT(μg / L); Chemical toxicity endpoint (EC i ) / EQS i Cases with values less than 1000 are also calculated; the chemical toxicity endpoint is the EC value of the target biotoxicity test for the chemical. 50 (LC) 50 or IGC 50 ), as detailed below:
[0101] (Option A)
[0102] (Option B)
[0103] (Option C)
[0104] (Option D)
[0105] (Option E)
[0106] Where EF represents the extrapolation factor.
[0107] Step 6-2: Use whether the difference between the predicted value and the actual value of EBT-BEQ / n from similar biotoxicity tests exceeds one order of magnitude as the standard for evaluating the accuracy of the calculation method; specifically, EBT-BEQ / n refers to the EBTs calculated based on Option C above.
[0108] Step 6-3: If the difference between the predicted value and the true value of EBT-BEQ / n exceeds one order of magnitude, calculate EBT-logBEQ using logBEQ, perform statistical analysis on EBT-BEQ and EBT-logBEQ, calculate EBT-BEQ using logBEQ based on the linear relationship between the two, and evaluate the accuracy of the calculation results according to the rules in Step 6-2.
[0109] EBT-BEQ is calculated using logBEQ based on the linear relationship between EBT-BEQ and EBT-logBEQ, as follows:
[0110] Linear fitting was performed on the true values of EBT-BEQ and EBT-logBEQ for the target biotoxicity test. The significance of the regression analysis results was tested by regression error analysis. Based on the linear analysis results, the mathematical relationship between EBT-BEQ and logBEQ was derived, and EBT-BEQ based on logBEQ was calculated. EBT-BEQ is the EBTs calculated based on Option D, and EBT-logBEQ is the EBTs obtained based on Option B.
[0111] Example 1:
[0112] like Figure 1 As shown, a specific embodiment of the method proposed in this invention includes the following steps:
[0113] S1. Establish the 96-hour median lethal concentration (LC50) of bighead carp. 50 The chemical toxicity characteristics database was used for data preprocessing.
[0114] S11. Collect and compile the CAS, CID, and 96-hour median lethal concentration (LC50) of chemicals from literature and databases (CHRIP, PubChem, etc.). 50 database;
[0115] S12. Based on the CAS and CID of the chemical, obtain the corresponding SMILES code through Pubchem, and use the SMILES as input to calculate the 2D molecular descriptor of the chemical as a candidate input feature.
[0116] S13. Remove missing values, outliers, and invalid values from the obtained molecular descriptors, eliminate descriptor columns with zero variance, and perform standardization to remove the influence of dimensions.
[0117] S2. Screen feature variables for model building, and establish and optimize the chemical target toxicity endpoint prediction model;
[0118] S21. Perform correlation analysis, retain the molecular descriptors with the largest variance after clustering if the correlation is ≥0.95, and use the F_regression method to screen candidate input features.
[0119] S22. Divide the dataset into a test set and a validation set in a 4:1 ratio, and perform internal validation using five-fold cross-validation. Using chemical molecule descriptors as input, predict the LC of the target biotoxicity test using a regression model. 50 The regression models include random forest, extreme gradient boosting tree, support vector regression, K-nearest neighbor algorithm, voting ensemble learning, stacking ensemble learning, Bagging ensemble learning, AdaBoost ensemble learning and multiple linear regression;
[0120] S23, Using R 2 MAE, RMSE, and MAPE characterize the model's predictive performance. Williams plots represent the application domain. The model's generalization ability is evaluated by assessing performance changes before and after outlier removal. Introducing application domain removal reduces the MAE and RMSE of the prediction model, but R... 2 The changes are not significant, indicating that the model has good generalization ability;
[0121] S24. Based on the evaluation parameters, determine the 96h LC of bighead carp. 50 The best predictive model is the Voting ensemble regression model (R²). 2 =0.666).
[0122] S3. Establish a database of chemical toxicity action patterns and preprocess the data;
[0123] S31. Collect and organize the CAS, CID, SID and SMILES structural information of organic chemicals from literature and databases (TOX21, PubChem, EBT dataset);
[0124] S32. Based on the chemical SMILES structure, the chemicals are classified into four categories using the Verhaar scheme module of Toxtree software: Class 1 (inert), Class 2 (weakly inert), Class 3 (reactive), and Class 4 (specific).
[0125] S33. Remove missing values, outliers, and invalid values from the dataset, eliminate descriptor columns with zero variance, and perform standardization. The preprocessed molecular descriptor dataset is used as a candidate input feature.
[0126] S4. Select feature variables for model building, and establish and optimize the chemical toxicity action mode classification model;
[0127] S41. Perform principal component analysis on the candidate input features and obtain the contribution of each principal component. Select the top 50 principal components as input features.
[0128] S42. Divide the dataset into a test set and a validation set in a 4:1 ratio. Take the screened molecular descriptors as input and the chemical toxicity mode information as output. Use classification algorithms to predict the toxicity mode of chemicals. Classification algorithms include: decision tree, random forest, support vector machine, Gaussian Naive Bayes model, logistic regression model and voting method.
[0129] S43. The model performance is characterized by AUC and confusion matrix, and the results are measured by Accuracy, Precision, Recall and F1 Score.
[0130] S44. Analyze the effectiveness of the four-category toxicity model classification based on inert chemicals, weakly inert chemicals, reactive chemicals, and chemicals with specific toxicity patterns of action. Figure 2 Based on the classification results, the scheme was adjusted, that is, inert chemicals and weakly inert chemicals were merged into non-specific chemicals, and chemicals were divided into three categories in total.
[0131] S5. Summarize the biotoxicity test results and toxic action modes of the target chemicals, and screen corresponding reference compounds;
[0132] S51. Collect the CAS of chemicals from the dataset (EBT-EQS) and calculate the SMILES structure of the chemicals based on PubChem.
[0133] S52. Based on the optimal target toxicity test endpoint prediction and toxicity mode classification model obtained from S2 and S4, supplement the calculation of the 96h LC of bighead carp for chemicals in the EBT-EQS dataset. 50 Test results and toxicity patterns;
[0134] S53. For the screening of reference compounds, priority should be given to the chemical with the highest toxicity (lowest toxicity endpoint value) among the chemicals in the specific toxicity category. If the toxicity endpoint value of the specific toxicity chemical is unavailable, the chemical with the highest toxicity among the reactive toxicity categories should be selected as the reference compound. If only the toxicity endpoint data of the non-specific toxicity category is available, the chemical with the highest toxicity among them should be directly selected as the reference compound. Based on this rule, the 96h LC50 of bighead carp was determined. 50 The reference compound was chlorpyrifos, such as... Figure 3 As shown.
[0135] S6. Calculate the corresponding biotoxicity effect threshold;
[0136] S61. Using the EBTs calculation method proposed by EBT-EQS, calculate the 96h LC of bighead carp supplemented by experimental data and predicted values. 50 The results of the EBTs are shown in Table 1.
[0137] Table 1. 96h LC of Bighead Carp 50 EBTs calculation results
[0138]
[0139] S62. Use whether the difference between EBT-BEQ / n in similar biotoxicity tests exceeds one order of magnitude as the standard for evaluating the predicted value of EBTs, and evaluate the accuracy of the calculation results.
[0140] S63, Bighead Carp 96h LC 50 The predicted value of EBT-BEQ / n differs from the actual value by more than an order of magnitude. EBT-logBEQ is calculated using logBEQ, and linear regression analysis is performed on EBT-BEQ and EBT-logBEQ. Based on the mathematical relationship between the two, the calculation process of EBT-BEQ is optimized using logBEQ. Comprehensive analysis yields the 96h LC value of the bighead carp. 50 The EBT-BEQs based on toxicity prediction values were 34.3 μg / L.
[0141] Example 2:
[0142] S1. Establish the 40-hour half-maximal growth inhibitory concentration (ICC) of Tetrahymena piriformis. 50 The chemical toxicity characteristics database was used for data preprocessing.
[0143] S11. Collect and organize the CAS, CID, and 40-hour median growth inhibitory concentration (IGC) of Tetrahymena piriformis from literature and databases (CHRIP, PubChem, etc.). 50 Toxicity database;
[0144] S12. Based on the CAS and CID of the chemical, obtain the corresponding SMILES code through Pubchem, and use the SMILES as input to calculate the 2D molecular descriptor of the chemical as a candidate input feature.
[0145] S13. Remove missing values, outliers, and invalid values from the obtained molecular descriptors, eliminate descriptor columns with zero variance, and perform standardization to remove the influence of dimensions.
[0146] S2. Screen feature variables for model building, and establish and optimize the chemical target toxicity endpoint prediction model;
[0147] S21. Perform correlation analysis, retain the molecular descriptors with the largest variance after clustering if the correlation is ≥0.95, and use the F_regression method to screen candidate input features.
[0148] S22. Divide the dataset into a test set and a validation set in a 4:1 ratio, and perform internal validation using five-fold cross-validation. Using chemical molecule descriptors as input, predict the IGC of the target biotoxicity test using a regression model. 50 The regression models include random forest, extreme gradient boosting tree, support vector regression, K-nearest neighbor algorithm, voting ensemble learning, stacking ensemble learning, Bagging ensemble learning, AdaBoost ensemble learning and multiple linear regression;
[0149] S23, Using R 2 MAE, RMSE, and MAPE characterize the model's predictive performance. Williams plots represent the application domain. The model's generalization ability is evaluated by assessing performance changes before and after outlier removal. Introducing application domain removal reduces the MAE and RMSE of the prediction model, but R... 2 The changes are not significant, indicating that the model has good generalization ability;
[0150] S24. Based on the evaluation parameters, determine the 40h IGC of Tetrahymena piriformis. 50 The best predictive model is the AdaBoost ensemble regression model (R²). 2 =0.783).
[0151] S3. Establish a database of chemical toxicity action patterns and preprocess the data;
[0152] S31. Collect and organize the CAS, CID, SID and SMILES structural information of organic chemicals from literature and databases (TOX21, PubChem, EBT dataset);
[0153] S32. Based on the chemical SMILES structure, the chemicals are classified into four categories using the Verhaar scheme module of Toxtree software: Class 1 (inert), Class 2 (weakly inert), Class 3 (reactive), and Class 4 (specific).
[0154] S33. Remove missing values, outliers, and invalid values from the dataset, eliminate descriptor columns with zero variance, and perform standardization. The preprocessed molecular descriptor dataset is used as a candidate input feature.
[0155] S4. Select feature variables for model building, and establish and optimize the chemical toxicity action mode classification model;
[0156] S41. Perform principal component analysis on the candidate input features and obtain the contribution of each principal component. Select the top 50 principal components as input features.
[0157] S42. Divide the dataset into a test set and a validation set in a 4:1 ratio. Take the screened molecular descriptors as input and the chemical toxicity mode information as output. Use classification algorithms to predict the toxicity mode of chemicals. Classification algorithms include: decision tree, random forest, support vector machine, Gaussian Naive Bayes model, logistic regression model and voting method.
[0158] S43. The model performance is characterized by AUC and confusion matrix, and the results are measured by Accuracy, Precision, Recall and F1 Score.
[0159] S44. Analyze the effectiveness of the four-category toxicity model classification based on inert chemicals, weakly inert chemicals, reactive chemicals, and chemicals with specific toxicity patterns of action. Figure 2 Based on the classification results, the scheme was adjusted, that is, inert chemicals and weakly inert chemicals were merged into non-specific chemicals, and chemicals were divided into three categories in total.
[0160] S5. Summarize the biotoxicity test results and toxic action modes of the target chemicals, and screen corresponding reference compounds;
[0161] S51. Collect the CAS of chemicals from the dataset (EBT-EQS) and calculate the SMILES structure of the chemicals based on PubChem.
[0162] S52. Based on the optimal target toxicity test endpoint prediction and toxicity mode classification model obtained from S2 and S4, supplement the calculation of Tetrahymena piriformis 40h IGC for chemicals in the EBT-EQS dataset. 50 Test results and toxicity patterns;
[0163] S53. According to the reference compound screening rules, Tetrahymena piriformis 40h IGC 50 The reference compound is 4-nonylphenol, such as Figure 3 As shown.
[0164] S6. Calculate the corresponding biotoxicity effect threshold;
[0165] S61. Using the EBTs calculation method proposed by EBT-EQS, calculate the 40h IGC of Tetrahymena piriformis supplemented by experimental data and predicted values. 50 The results of the EBTs are shown in Table 2.
[0166] Table 2 Tetrahymena piriformis 40h IGC 50EBTs calculation results
[0167]
[0168] S62. Analyze the accuracy of the calculation results according to the evaluation criteria;
[0169] S63, Tetrahymena piriformis 40h IGC 50 The predicted value of EBT-BEQ / n differs from the actual value by more than an order of magnitude. EBT-logBEQ is calculated using logBEQ, and linear regression analysis is performed on EBT-BEQ and EBT-logBEQ. Based on the mathematical relationship between the two, the calculation process of EBT-BEQ is optimized using logBEQ. Comprehensive analysis yields the following results for Tetrahymena piriformis 40h IGC. 50 The EBT-BEQs based on toxicity prediction values were 11.8 μg / L.
[0170] This embodiment provides a framework and approach for calculating the threshold of aquatic organism toxicity effects based on machine learning. This method can supplement data on the acute toxic effects of chemicals on aquatic organisms, providing a reference for the identification and management of aquatic health risks.
[0171] In its specific implementation, this application provides a computer storage medium and a corresponding data processing unit. The computer storage medium is capable of storing a computer program, which, when executed by the data processing unit, can run the invention's content regarding a machine learning-based method for calculating the threshold of toxic effects on aquatic organisms, as well as some or all of the steps in various embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.
[0172] Those skilled in the art will clearly understand that the technical solutions in the embodiments of the present invention can be implemented using computer programs and their corresponding general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of computer programs, i.e., software products. These computer program software products can be stored in a storage medium and include several instructions to cause a device containing a data processing unit (which may be a personal computer, server, microcontroller, MUU, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.
[0173] This invention provides a method for calculating the threshold of toxic effects on aquatic organisms based on machine learning. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.
Claims
1. A method for calculating the threshold of toxic effects on aquatic organisms based on machine learning, characterized in that, Includes the following steps: Step 1: Construct a database of chemical target toxicity endpoint features. The chemical target toxicity endpoint features include chemical molecule descriptor information, corresponding target biotoxicity test results information, and preprocess the feature data. Step 2: Establish and optimize the chemical target toxicity endpoint prediction model. Based on the database established in Step 1, use the feature screening method to select chemical molecular descriptors as input; construct a regression model, use the screened molecular descriptors as input, predict the endpoint of the target biotoxicity test, and construct the target toxicity endpoint prediction model. Step 3: Construct a chemical toxicity mode database. The database includes chemical molecule descriptor information and corresponding toxicity mode characteristics. The data is then preprocessed. Step 4: Establish a chemical toxicity mode prediction model. Based on the database established in Step 3, use principal component analysis to screen chemical molecule descriptors as input; input the screened molecule descriptors into the classification model to predict their toxicity mode and construct a toxicity mode classification model. Step 5: Based on the target toxicity endpoint prediction model and toxicity mode prediction model obtained in Steps 2 and 4, supplement the compound toxicity characteristic parameters and screen reference compounds for the target toxicity test type of chemical mixtures. Step 6: Based on the reference compounds screened out in Step 5, calculate the toxicity effect threshold of the target toxicity endpoint for aquatic organisms.
2. The method for calculating the threshold of toxic effects on aquatic organisms based on machine learning according to claim 1, characterized in that, The construction of the chemical target toxicity endpoint characterization database described in step 1 includes: Step 1-1: Collect and organize data on the endpoints of chemical toxicity tests for aquatic organisms from literature and databases; Steps 1-2: Based on the SMILES structure of the chemical, calculate the corresponding 2D molecular descriptor, remove missing values, outliers and invalid values, eliminate descriptor columns with variance of 0, and perform standardization.
3. The method for calculating the threshold of toxic effects on aquatic organisms based on machine learning according to claim 2, characterized in that, Step 2, which involves establishing and optimizing the chemical target toxicity endpoint prediction model, includes: Step 2-1: Calculate the distance between molecular descriptor vectors, cluster molecular descriptors with correlation ≥ 0.95, retain only the molecular descriptor with the largest variance in each cluster, and use a filtering method to select molecular descriptors as input. Step 2-2: Divide the dataset into a test set and a validation set according to the proportion, and perform internal validation using five-fold cross-validation. Take the molecular descriptors screened out in Step 2-1 as input and the endpoint data of the toxicity test as output. Use regression algorithms including random forest, extreme gradient boosting tree, support vector regression, K-nearest neighbor algorithm, voting ensemble learning, stacking ensemble learning, Bagging ensemble learning, AdaBoost ensemble learning and multiple linear regression to train and build the model. Steps 2-3: Use mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). 2 Mean absolute percentage error (MAPE) characterizes the model's predictive performance. Steps 2-4: Draw Williams diagrams based on standardized residuals and lever distances to represent the application domain; Steps 2-5, with 𝑅 2 >0.500 and MAPE<100% are the criteria for judging the model performance as qualified. The best predictive model for the target biotoxicity test endpoint is determined based on the model performance.
4. The method for calculating the threshold of toxic effects on aquatic organisms based on machine learning according to claim 3, characterized in that, Step 3, which involves constructing a database of chemical toxicity patterns, includes: Step 3-1: Collect and organize the CAS, CID, SID, and SMILES structural information of organic chemicals from literature and databases; Step 3-2: Based on the SMILES structure of chemicals, chemicals are classified into four categories: inert chemicals (Class 1), weakly inert chemicals (Class 2), reactive chemicals (Class 3), and specific chemicals (Class 4). Step 3-3: Remove missing values, outliers, and invalid values, eliminate descriptor columns with a variance of 0, and perform standardization.
5. The method for calculating the threshold of toxic effects on aquatic organisms based on machine learning according to claim 4, characterized in that, Step 4, which involves establishing a predictive model for the toxic effects of chemicals, includes: Step 4-1: Perform principal component analysis on the input variables and obtain the contribution of each principal component. Select the top 50 principal components as input features. Step 4-2: Divide the dataset into a test set and a validation set according to the proportion. Take the input features screened out in Step 4-1 as input and the chemical toxicity mode information as output. Use classification models including decision tree, random forest, support vector machine, Gaussian Naive Bayes model, logistic regression model and voting method for training. Step 4-3: Use the area under the ROC curve (AUC) and confusion matrix, and select accuracy, precision, recall, and F1 score to measure the results. Step 4-4: Determine the best classification scheme based on the performance of the classification model. Analyze the performance of the four-category toxicity action pattern classification model based on the evaluation indicators of the classification model, including inert chemicals, weakly inert chemicals, reactive chemicals, and chemicals with specific toxicity action patterns. Adjust the scheme according to the classification performance and continue to use the adjusted scheme for model training to evaluate the model performance.
6. The method for calculating the threshold of toxic effects on aquatic organisms based on machine learning according to claim 5, characterized in that, The adjustment of the scheme based on the classification effect described in step 4-4 includes: analyzing the accuracy of the constructed classification model, merging Class 1 of inert chemicals and Class 2 of weakly inert chemicals into a non-specific chemical category, and classifying chemicals into three categories in total: non-specific chemicals, reactive chemicals, and specific chemicals.
7. The method for calculating the threshold of toxic effects on aquatic organisms based on machine learning according to claim 6, characterized in that, The reference compounds for screening chemical mixtures for the target toxicity test type described in step 5 include: Step 5-1: Collect the CAS of the substances contained in the chemical mixture and calculate the SMILES structural formula of the chemical. Step 5-2: Based on the chemical target toxicity endpoint prediction model obtained in Step 2, expand the data for the biotoxicity test endpoints of substances in chemical mixtures. Step 5-3: Based on the chemical toxicity mode prediction model obtained in Step 4, the toxicity mode information of substances in the chemical mixture is expanded. Step 5-4: For the target biotoxicity test, prioritize the chemical with the lowest toxicity endpoint value among the specific category chemicals as the reference compound; if the toxicity endpoint value of the specific toxic chemical is unavailable, select the chemical with the lowest toxicity endpoint value among the reactive category chemicals as the reference compound; if only the toxicity endpoint data of the non-specific category chemicals is available, directly select the chemical with the lowest toxicity endpoint value among them as the reference compound.
8. The method for calculating the threshold of toxic effects on aquatic organisms based on machine learning according to claim 7, characterized in that, Step 6, which involves calculating the toxicity effect threshold of the target toxicity endpoint for aquatic organisms, includes: Step 6-1: Using the EBTs calculation method proposed by EBT-EQS, calculate the effect-based trigger values (EBTs) for the target biotoxicity test of the chemical mixture. Step 6-2: Use whether the difference between the predicted value and the actual value of EBT-BEQ / n in the same type of biotoxicity test exceeds one order of magnitude as the standard for evaluating the predicted value of EBTs, and evaluate the accuracy of the calculation method. Step 6-3: If the difference between the predicted value and the true value of EBT-BEQ / n exceeds one order of magnitude, calculate EBT-logBEQ using logBEQ, perform statistical analysis on EBT-BEQ and EBT-logBEQ, calculate EBT-BEQ using logBEQ based on the linear relationship between the two, and evaluate the accuracy of the calculation results according to the rules in Step 6-2.
9. The method for calculating the threshold of toxic effects on aquatic organisms based on machine learning according to claim 8, characterized in that, The EBT-BEQ / n mentioned in step 6-2 refers to the calculation of EBT based on the following formula: ， in, Let represent the bioeffect equivalent of the i-th chemical, and n represent the quantity of the chemical.
10. The method for calculating the threshold of toxic effects on aquatic organisms based on machine learning according to claim 9, characterized in that, Step 6-3, which calculates EBT-BEQ using logBEQ based on the linear relationship between EBT-BEQ and EBT-logBEQ, includes: Linear fitting was performed on the true values of the corresponding biotoxicity tests EBT-BEQ and EBT-logBEQ. The significance of the regression analysis results was tested by regression error analysis. Based on the linear analysis results, the mathematical relationship between EBT-BEQ and logBEQ was obtained, and EBT-BEQ based on logBEQ was calculated.