Method for predicting toxicity of PFAS on microalgae based on density functional and step-by-step recursive feature elimination

Abstract

The invention discloses a method for predicting the toxicity of PFAS to microalgae based on density functional and step-by-step recursive feature elimination, and the method comprises the following steps: (1) obtaining two microalgae toxicity experiment data sets based on an ECOTOX database, and respectively using the two microalgae toxicity experiment data sets for internal training and external verification; (2) performing density functional theory (DFT) calculation on the basis of a B3LYP / 6-31 + G (d, p) group by adopting Gaussian 09 software, and optimizing the geometrical shape of the PFAS to obtain an optimal conformation; (3) calculating and analyzing by using Multiwfn 3.7 software to obtain a chemical descriptor of the molecule; (4) performing feature selection from the descriptors by using a step-by-step recursive feature elimination (RFE) method taking a ridge regression model as a base model; (5) modeling the toxicity index by using a multiple linear regression model; and (6) verifying the constructed model by using an external data set.

Description

Technical Field

[0001] The present invention relates to the field of environmental science, and in particular to predicting the toxic effects of per- and polyfluoroalkyl substances (PFAS) on microalgae by using quantitative structure-activity relationship (QSAR) and molecular simulation technology. Background Art

[0002] Per- and polyfluoroalkyl substances (PFAS) have good chemical stability, thermal stability and high surface activity and are widely used in more than 200 application fields such as fire extinguishing foam, food packaging, non-stick cookware, surfactants, electroplating additives, etc. More than one million tons of PFAS are produced each year worldwide, and thousands of PFASs eventually enter the environment through surface runoff, wastewater discharge, product degradation, etc. Due to their persistence and bioaccumulation, some PFASs have been detected in humans and animals with a half-life of about 3.5 years, and some long-chain PFASs have been shown to cause liver toxicity, accelerate skin aging, damage lipid metabolism, immune and endocrine systems, and multi-organ tumors. As an emerging pollutant, PFASs are receiving more and more attention.

[0003] As primary producers, microalgae play an important role in maintaining the balance of aquatic ecosystems. Once PFAS enters aquatic ecosystems, it may change the nutrient cycle process and have adverse effects on algae and even higher trophic organisms in the food chain. At present, there are relatively few data and mechanisms on the impact on algae. However, most of the existing toxicity data focus on PFOA and limited alternatives. PFAS is a very large family of compounds with more than 8,000 known structures, and there are still a large number of PFAS toxic effects that have not been studied. In 2009, the Stockholm Convention on Persistent Organic Pollutants initiated the regulation of the use of perfluorooctane sulfonic acid (PFOS), its salts and related substances. Increasing restrictions and public concern about these chemicals have led to the emergence of more and more alternatives. The huge market demand has driven manufacturers to seek more and more alternatives. Inevitably, these alternatives have been observed in the environment and even in biological matrices. However, although the experimental measurement data is reliable, the time cost is high and the economic cost is very expensive, so it is particularly important to find a convenient and low-cost toxicity prediction method. Summary of the invention

[0004] In response to the relative lack of environmental ecotoxicity data for the large family of PFAS compounds, we developed an interpretable multivariate linear regression equation to predict the toxicity of PFAS to microalgae, which was well validated in per- and polyfluoroalkyl compounds. This new approach outperforms traditional machine learning methods in terms of interpretability, thus providing a convenient tool for cytotoxicity prediction.

[0005] The technical solution of the present invention is a method for predicting the toxicity of PFAS to microalgae based on density functional theory and stepwise recursive feature elimination, comprising the following steps:

[0006] (1) Two sets of PFAS toxicity data on acute experiments on microalgae were obtained based on the ECOTOX database, with the half inhibitory concentration IC50 as the toxicity indicator;

[0007] (2) The chemical structures of PFAS were collected from the PubChem website ( https: / / pubchem.ncbi.nlm.nih.gov / ) to ensure the accuracy of the basis for descriptor calculation;

[0008] (3) Density functional theory (DFT) calculations were performed using Gaussian 09 software based on the B3LYP / 6-31+G(d,p) basis set to optimize the geometry of PFAS, obtain the minimum energy conformation, and determine the structural parameters;

[0009] (4) In the process of in-depth molecular chemical descriptor calculation and analysis using Multiwfn 3.7 software, we systematically counted the strongly correlated pairs with an absolute value of correlation coefficient exceeding 0.7 among all descriptors. In order to optimize the descriptor set and reduce redundant information, we implemented a sophisticated preprocessing step: first, each descriptor was sorted according to the number of absolute values ​​of correlation coefficients between it and other descriptors greater than 0.7; then, following the principle of stepwise elimination, descriptors with the most extensive high correlation were deleted one by one until the correlation between any two descriptors in the entire descriptor set no longer exceeded the threshold of 0.7;

[0010] (5) Use Python and scikit-learn to perform recursive feature elimination (RFE) to rank the selected descriptors, use the ridge regression model as the base model, and introduce L2 regularization to penalize larger regression coefficients to avoid overfitting problems;

[0011] (6) Multiple linear regression equations were used for modeling, and four key indicators were used for measurement: coefficient of determination (R 2 ), p-value of F statistic, variance inflation factor (VIF), and Durbin-Watson (DW) statistic;

[0012] (7) Use external datasets to validate the model.

[0013] Furthermore, the two groups of PFAS toxicity data on acute experiments on microalgae in step (1) are: 72h-IC50 data of 8 fluorinated acids on P. subcapitata and IC50 data of 6 perfluorocarboxylic acids on P. subcapitata, which are used for model training and verification respectively, and are both acute toxicity data of fluorinated acids on microalgae.

[0014] Furthermore, in step (2), the chemical structures of PFAS were collected from the PubChem website, with 3D structures being prioritized. If there was no 3D structure, a 2D structure was selected.

[0015] Furthermore, the L2 regularization coefficient in step (5) is set to 1.0.

[0016] Furthermore, the indicator calculation formula and evaluation principle of step (6) are as follows:

[0017] 1) Coefficient of determination (R 2 )

[0018] Coefficient of determination (R 2 ) measures the proportion of the variation explained by the model to the total variation. Its value ranges from 0 to 1. The closer it is to 1, the better the model fit.

[0019]

[0020] Among them, y i is the observed value; is the predicted value; The mean of the observations;

[0021] 2) p-value of F statistic

[0022] The p-value of the F statistic is used to test the significance of the model as a whole. If the p-value is less than the set significance level (usually 0.05), the model is considered significant.

[0023]

[0024] Where TSS is the total sum of squares; RSS is the residual sum of squares; n is the number of observations; p is the number of independent variables;

[0025] 3) Variance Inflation Factor (VIF)

[0026] Variance inflation factor (VIF) is used to detect multicollinearity problems in the model. VIF values ​​greater than 5 or 10 are usually considered to be serious multicollinearity.

[0027]

[0028] in, is the coefficient of determination when the jth independent variable is regressed on the remaining independent variables;

[0029] 4) DW statistics

[0030] The DW statistic is used to detect autocorrelation in the residual sequence. The ideal DW value is close to 2, which means that there is no obvious autocorrelation in the residual sequence.

[0031]

[0032] Among them, e i is the residual of the ith observation;

[0033] Together, these metrics provide important information about the quality of the model, helping us judge its predictive power and reliability.

[0034] The present invention has the following advantages:

[0035] 1. This method predicts the toxicity of PFASs through mathematical models, which is more efficient and faster than traditional laboratory testing methods. By calculating molecular descriptors, the potential toxicity of a large number of PFASs can be predicted in a short period of time without the need for time-consuming and labor-intensive experimental testing one by one.

[0036] 2. The model has good predictive ability and stability. The model has been rigorously validated using internal and external data sets. 2 It reaches 0.84 and 0.99 in the training set and validation set respectively.

[0037] 3. Compared with black box models such as machine learning, the multivariate linear regression model used in this invention is a highly interpretable model, which is convenient for researchers to understand the logic behind the model prediction results. Through the selected molecular descriptors (ODI_LUMO, Mol_Size_2, ODI_HOMO), it can be directly linked to the physicochemical properties of PFASs, thereby explaining their toxic mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is the heat map of the correlation between the eigenvalue and IC50;

[0039] Figure 2 The relationship between feature descriptors and IC50 in the model;

[0040] Figure 3 (a) Correlation between the model prediction values ​​and actual observation values ​​in the training set and (b) validation set. DETAILED DESCRIPTION

[0041] The present invention is further described below in conjunction with the accompanying drawings and specific implementations. It should be emphasized that the following description is only for illustrative purposes and is not intended to limit the scope of use and application of the present invention.

[0042] The present invention selects and constructs a model according to the model performance indicators under different numbers of eigenvalues ​​in Table 1.

[0043] Table 1 Model performance indicators under different numbers of eigenvalues

[0044]

[0045] When the number of eigenvalues ​​increases from 1 to 5, R2 increases from 0.628 to 0.973. However, when the number of eigenvalues ​​is 4 or 5, the significance test p>0.05, and when the number of eigenvalues ​​is 5, VIF increases significantly, which may also cause collinearity problems. When the number of eigenvalues ​​is 3, p<0.05, and VIF is small, which conforms to the principle of building a model with the least number of features, so the three descriptors ODI_LUMO, Mol_Size_2, and ODI_HOMO are finally selected for model construction.

[0046] Figure 3 (a) Training set and validation set (b) Correlation between model prediction value and actual observation value, training set model R 2 =0.84, validation set model R 2 =0.99.

[0047] A method for predicting the toxicity of PFAS to microalgae based on density functional theory and stepwise recursive feature elimination includes the following steps:

[0048] (1) Based on the ECOTOX database, two sets of toxicity experimental data sets on microalgae were obtained, which were used for internal training and external validation, respectively. The half inhibitory concentration IC50 was selected as the toxicity indicator.

[0049] (2) Density functional theory (DFT) calculations were performed using Gaussian 09 software based on the B3LYP / 6-31+G(d,p) basis set to optimize the geometry and minimum energy conformation of PFAS and determine the structural parameters;

[0050] (3) The chemical descriptors of the molecules were calculated and analyzed using Multiwfn 3.7 software;

[0051] (4) Calculate the correlation number of molecular descriptors in the same class and establish a descriptor correlation matrix. Select the descriptor with the most correlation descriptors and delete it from the correlation matrix to initially simplify the feature descriptors.

[0052] Then, feature selection is performed from the descriptors using the stepwise recursive feature elimination (RFE) method with the ridge regression model as the base model;

[0053] (5) Modeling of toxicity indicators using a multivariate linear regression model;

[0054] (6) Use external data sets to verify the constructed model.

[0055] In the embodiment, the modeling analysis is performed using the Person correlation matrix and step-by-step recursive feature elimination, including the following steps:

[0056] First, data import: The PFAS molecules involved in the toxicity data set obtained from ECOTOX are used to perform density functional theory (DFT) calculations based on the B3LYP / 6-31+G(d,p) basis set using Gaussian 09 software. Water molecules are added to the solvation to obtain the minimum energy conformation and determine the structural parameters. The output result is a .out file.

[0057] Secondly, the chemical descriptors of the molecules were calculated and analyzed using Multiwfn 3.7 software, in which the .out files were converted into .fchk files for batch processing.

[0058] Then, the calculated chemical descriptors of the same type were subjected to correlation matrix analysis, and the descriptors with absolute values ​​of correlation coefficients greater than 0.7 were calculated. The descriptors with the most correlated descriptors were selected and deleted from the correlation matrix. Through this operation, the feature descriptors were initially streamlined.

[0059] Again, recursive feature elimination (RFE) is used to further rank the selected descriptors. Ridge regression model is used as the base model, and L2 regularization is introduced (regularization coefficient α is set to 1.0) to penalize larger regression coefficients to avoid overfitting problems. RFE is performed using Python and scikit-learn.

[0060] The multiple linear regression equation was used for modeling.

[0061] Finally, we used an external data set and compared the IC50 values ​​measured by the external data experiment with the IC50 values ​​predicted by this model.

Claims

1. A method for predicting the toxicity of PFAS to microalgae based on density functional theory and stepwise recursive feature elimination, characterized in that: The following steps are involved: (1) Two sets of PFAS toxicity data on acute experiments on microalgae were obtained based on the ECOTOX database, with the half inhibitory concentration IC50 as the toxicity indicator; (2) The chemical structures of PFAS were collected from the PubChem website ( https: / / pubchem.ncbi.nlm.nih.gov / ) to ensure the accuracy of the basis for descriptor calculation; (3) Density functional theory (DFT) calculations were performed using Gaussian 09 software based on the B3LYP / 6-31+G(d,p) basis set to optimize the geometry of PFAS, obtain the minimum energy conformation, and determine the structural parameters; (4) The chemical descriptors of the molecules were calculated and analyzed using Multiwfn 3.7 software. The descriptors with absolute values ​​of correlation coefficients greater than 0.7 were statistically analyzed and optimized until the correlation between any two descriptors in the entire descriptor set no longer exceeded the threshold of 0.7; (5) Recursive feature elimination (RFE) was performed using Python and scikit-learn to rank the selected descriptors, the ridge regression model was used as the base model, and L2 regularization was introduced to penalize larger regression coefficients to avoid overfitting problems; (6) The multivariate linear regression equation was used for modeling, and four key indicators were used for measurement: coefficient of determination, p-value of F statistic, variance inflation factor, and Durbin-Watson statistic; (7) Use external datasets to validate the model.

2. The prediction method according to claim 1, characterized in that: The two groups of PFAS toxicity data on acute experiments on microalgae in step (1) are: 72h-IC50 data of 8 fluorinated acids on P. subcapitata and IC50 data of 6 perfluorocarboxylic acids on P. subcapitata, which are used for model training and verification respectively, and are both acute toxicity data of fluorinated acids on microalgae.

3. The prediction method according to claim 1, characterized in that: In the step (2), the chemical structures of PFAS are collected from the PubChem website, with 3D structures being prioritized. If there is no 3D structure, a 2D structure is selected.

4. The prediction method according to claim 1, characterized in that: The L2 regularization coefficient in step (5) is set to 1.

0.

5. The prediction method according to claim 1, characterized in that: The indicator calculation formula and evaluation principle of step (6) are as follows: 1) Determination coefficient R 2 Coefficient of determination R 2 It measures the proportion of the variation explained by the model to the total variation. Its value ranges from 0 to 1. The closer it is to 1, the better the model fit. Among them, y i is the observed value; is the predicted value; The mean of the observations; 2) p-value of F statistic The p-value of the F statistic is used to test the significance of the model as a whole. If the p-value is less than the set significance level, usually 0.05, the model is considered significant. Where TSS is the total sum of squares; RSS is the residual sum of squares; n is the number of observations; p is the number of independent variables; 3) Variance Inflation Factor (VIF) Variance inflation factor (VIF) is used to detect multicollinearity problems in the model. VIF values ​​greater than 5 or 10 are usually considered to be serious multicollinearity. in, is the coefficient of determination when the jth independent variable is regressed on the remaining independent variables; 4) DW statistics The DW statistic is used to detect autocorrelation in the residual sequence. The ideal DW value is close to 2, which means that there is no obvious autocorrelation in the residual sequence. Among them, e i is the residual of the ith observation; Together, these metrics provide important information about the quality of the model, helping us judge its predictive power and reliability.

Images