CN111768813A - Prediction method of organic PDMS membrane-water partition coefficient based on quantitative structure-activity relationship model based on SW-SVM algorithm - Google Patents

Abstract

The invention discloses a method for organic PDMS membrane-water distribution coefficient, which is characterized in that a molecular descriptor is calculated through the molecular structure of the existing compound, a stepwise linear regression-support vector machine (SW-SVM) analysis combination method is adopted, a quantitative structure-property relation model is constructed, and the organic compound can be rapidly and efficiently predictedK _PDMS‑wA value; the method is simple and quick, has low cost, can save manpower, material resources and financial resources required by experimental tests, develops a nonlinear model with better generalization capability by using R language, and has good goodness of fit, robustness and prediction capability; the invention can effectively predict the PDMS film/water distribution coefficient of the organic compound in the application domain, fills the blank of data of other compounds, and is used for monitoring and passive sampling of environmental compoundsThe application of the device provides necessary basic data and has great significance.

Description

Quantitative structure-activity relationship model based on SW-SVM algorithm to predict organic matter PDMS membrane-water method of distribution coefficient

technical field

The invention relates to a method for organic matter PDMS membrane/water partition coefficient, in particular to a method for predicting organic matter PDMS membrane-water partition coefficient based on a quantitative structure-activity relationship model based on SW-SVM algorithm.

Background technique

Membrane passive sampling techniques are a widely accepted method for measuring freely dissolved concentrations of organic compounds and assessing their environmental exposure risk. PDMS (polydimethylsiloxane) membranes have become one of the most widely used passive sampling materials due to their good thermal stability and high affinity for hydrophobic compounds. Usually, the partition coefficient of organic matter between PDMS membrane and water (K _PDMS-w ) is an important parameter for evaluating the environmental behavior of compounds, and it is also the key to the successful application of passive samplers. The measurement of conventional experiments is time-consuming and labor-intensive, and it is difficult to meet the large and growing needs of environmental monitoring and management of organic pollutants. Therefore, it is particularly important to develop a simple and accurate theoretical prediction method for estimating the K _PDMS-w of organic matter.

Quantitative structure-property relationship (QSPR) refers to a computer modeling method that correlates the molecular structure of organic matter with its physicochemical properties, environmental behavior, and toxicological parameters, which can reduce or replace related experiments, make up for the lack of experimental data, and reduce experimental costs. At present, there are few research methods on K _PDMS-w for predicting organic matter, especially the research on nonlinear methods, and the number of substances studied in the existing model is relatively small, and the prediction accuracy of the model needs to be further improved. Considering that the environmental allocation behavior of various organic compounds is a complex process, the partition coefficient may involve some nonlinear relationships. Therefore, it is necessary to construct a K that covers a variety of compounds, has a clear algorithm, is easy to apply, and does not depend on experimental data. _PDMS-w nonlinear prediction model, and the model is validated and characterized according to OECD guidelines.

Considering the high dimensionality of feature descriptors, how to select the most useful subset of features from the original variables for modeling becomes increasingly important. In order to select more reasonable molecular descriptors for constructing the QSPR model, a stepwise linear regression screening method was adopted to achieve the purpose of variable dimensionality reduction. In order to establish a reliable QSPR model, the nonlinear algorithm used is the support vector machine (SVM) regression algorithm, which is not only simple and easy to implement, but also has good "robustness" and excellent generalization ability. Models can also be made reproducible by calling the R language package e1071 to set the number of seeds.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for predicting the organic PDMS membrane-water partition coefficient based on the quantitative structure-activity relationship model of the SW-SVM algorithm, which can quickly and effectively predict the K _PDMS-w value of the organic compound according to the molecular structure descriptor of the organic compound. , which is conducive to the risk assessment of pollutants, to decision makers and managers to formulate relevant standards for chemical emissions, and to providing new ideas for environmental governance.

The object of the present invention is achieved in this way: a method for predicting the organic matter PDMS membrane-water partition coefficient based on the quantitative structure-activity relationship model of the SW-SVM algorithm, is characterized in that, comprises the following steps:

Step 1) Data collection: The log K _PDMS-w value containing several organic compounds was collected by referring to the literature, and 1/5 of the obtained data set was extracted according to the size of its log K _PDMS-w value as the validation set data, and the rest were used as the validation set data. training set data;

Step 2) Descriptor calculation: use MM2 molecular mechanics to optimize the initial molecular structure of organic compounds, use alvaDesc 1.0.0 to obtain molecular structure descriptors of organic compounds, and screen out the final descriptors through stepwise linear regression after preprocessing;

Step 3) Model construction: take the final descriptor as the independent variable, the log K _PDMS-w of the organic matter PDMS-water partition coefficient as the dependent variable, use the support vector machine regression algorithm to establish the QSPR prediction model on the training set, and pass the k-fold cross-validation. The algorithm selects the optimal parameters to construct the QSPR model based on the optimal SW-SVM algorithm;

Step 4) Model verification: the model is verified, which is divided into two steps: a) goodness of fit and robustness evaluation of the model; b) application domain characterization and performance evaluation of the model; after passing the verification, go to step 5);

Step 5) Application domain characterization: the model application domain is characterized by the Williams diagram;

Step 6) Model application: use the model to predict the PDMS membrane/water partition coefficient of the unknown compound.

As a further limitation of the present invention, in the step 1), the organic compounds include polycyclic aromatic hydrocarbons, polychlorinated biphenyls, benzenes, pesticides, ethers, dioxins, esters, aliphatics, hydrocarbons, nitrogen Sulfur compounds.

As a further limitation of the present invention, in step 1), for the same substance in the collected organic compounds, the data that obviously deviates from the overall value are excluded, and the average value is taken to conduct model construction research.

As a further limitation of the present invention, the organic compounds in the training set described in step 1) are used for constructing the model for internal verification, and the organic compounds in the verification set are used for external verification of the model

As a further limitation of the present invention, in step 2), the preprocessing process includes removing constants, near constants, missing descriptors and descriptors with a correlation greater than 0.95.

As a further limitation of the present invention, in step 3), use the R language program package to build the QSPR model based on the optimal SW-SVM algorithm, specifically including the following process:

Step 3-1, first divide the entire data set into k sets, each set will be used as the test set in turn, and the remaining sets will be used as the training set. Repeat the training and testing k times to ensure that each set is used as the test set. verified once;

Step 3-2, calculate and compare the average cross-validation accuracy of k-fold training, and select a set of parameters with the highest cross-validation accuracy. This set of parameters (cost, gamma) will be used as the optimal value of k-fold cross-validation to apply In vector machine regression prediction, the penalty factor cost controls the relative proportion of model structural risk and empirical risk, and determines the superiority of the model. The gamma parameter determines the distribution of the data after mapping to the new feature space. The prediction model selects gamma as the diameter. To the basis kernel function, such as the formula g=1/2σ ² , where σ is the width parameter of the function, which controls the radial action range of the function;

Steps 3-3, apply the parameters to the model to construct the optimal model.

训练集均方根误差RMSE_tra以及训练集平均绝对误差MAE_tra。As a further limitation of the present invention, in the step 4, when the model is verified, its goodness of fit and robustness evaluation indicators are: the coefficient of determination of the degree of freedom correction

The training set root mean square error RMSE _tra and the training set mean absolute error MAE _tra .

As a further limitation of the present invention, step 5) specifically includes: using the Williams diagram of the leverage value _hi based on the standard residual δ to characterize the application domain of the model, when the absolute value of δ is greater than 3.0, the compound is an outlier, When the leverage value _hi is greater than the warning value h*, it indicates that the structure of this compound is significantly different from other compounds; _hi and h* are calculated by the following formulas:

hi = x _i ^T (X ^T X) ^-1 x _i

h*=3(p+1)/n

where x _i is the descriptor matrix of the ith compound; x _i ^T is the transpose matrix of x _i ; X is the descriptor matrix of all compounds; X ^T is the transpose matrix of X; (X ^T X) ^-1 is Inverse of matrix X ^T X; p is the number of variables in the model and n is the number of data points in the dataset.

Compared with the prior art, the beneficial effects of the present invention are:

1. According to the OECD guidelines on the construction and use of QSRR models, the established QSRR models have good fit, robustness and predictive ability;

2. The model has a wide range of applications, covering organic compounds of various structures, and can be used to predict the K _PDMS-w value of different compounds, providing basic data for global environmental behavior analysis and ecological risk assessment of organic compounds;

3. The model completely adopts the calculation method, which is different from the previous method relying on the experimental value, which can greatly reduce the experimental cost and obtain the K _PDMS-w value of the chemical more efficiently;

4. The present invention can quickly and effectively predict the PDMS membrane/water partition coefficients of various organic compounds. The method is low-cost, simple and fast, and can save a lot of manpower, material resources and financial resources. The establishment and verification of the QSRR model involved in the invention is strictly in accordance with the guidelines for the construction and use of the QSPR model stipulated by the OECD, which is accurate and reliable, can effectively obtain the K _PDMS-w value of the substance, and provides important basic data for chemical supervision work. It has important guiding significance for ecological risk assessment.

Description of drawings

FIG. 1 is a flow chart of the prediction method of the present invention.

FIG. 2 is a fitting diagram of the experimental value and the predicted value of the data set log K _PDMS-w of the present invention.

Figure 3 is a Williams diagram in the present invention.

Detailed ways

As shown in Figure 1, a method for predicting the organic PDMS membrane-water partition coefficient based on the quantitative structure-activity relationship model of the SW-SVM algorithm includes the following steps.

Step 1) The log K _PDMS-w value of 347 kinds of organic compounds was obtained by referring to the literature. For the same substance, the data that deviates significantly from the overall value were excluded, and the average value was taken to construct the model. The organic compounds include polycyclic aromatic hydrocarbons, polychlorinated Biphenyls, benzenes, pesticides, ethers, dioxins, esters, aliphatics, hydrocarbons, nitrogen-sulfur compounds; sort the obtained data sets according to their log K _PDMS-w values, and take the first 4/ 5 organic compounds are used as training set data, and the rest are used as validation set data. The training set data includes 277 and the validation set data includes 70. The organic compounds in the training set are used to build the model for internal verification, and the organic compounds in the validation set are used for Validation outside the model.

Step 2) Use the MM2 molecular mechanics method to optimize the initial molecular structure of organic compounds, use alvaDesc to obtain the molecular structure descriptors of organic compounds, remove constants, close to constants, deletions and descriptors with a correlation greater than 0.95, and then screen by stepwise linear regression out the final descriptor.

Step 3) Take the final descriptor screened out as the independent variable, the log K _PDMS-w of the organic matter PDMS/water partition coefficient as the dependent variable, and call the support vector machine regression algorithm of the R language e1071 package to establish a QSPR prediction model, and pass k The folding cross-validation algorithm selects the optimal parameters (cost, gamma), and uses the R language package e1071 to construct the QSPR model based on the optimal SW-SVM algorithm. The specific process is as follows:

Step 3-1, first divide the entire data set into k sets, each set will be used as a test set in turn, and the remaining sets will be used as a training set, so that the training and testing are repeated k times to ensure that each set is used as a test set. verified once;

Step 3-2, calculate and compare the average cross-validation accuracy of k-fold training, and select a set of parameters with the highest cross-validation accuracy. In machine regression prediction, the penalty factor cost controls the relative proportion of model structural risk and empirical risk, and determines the superiority of the model. The gamma parameter determines the distribution of the data after mapping to the new feature space. The prediction model selects gamma as the radial The basis kernel function, such as the formula g=1/2σ ² , where σ is the width parameter of the function, which controls the radial action range of the function;

Steps 3-3, apply the parameters to the model to construct the optimal model.

Step 4, the verification and characterization of the model is divided into two steps: 1) goodness of fit and robustness evaluation of the model; 2) application domain characterization and performance evaluation of the model;

训练集均方根误差RMSE_tra和训练集平均绝对误差MAE_tra表征，在本实施例中自由度校正的决定系数

训练集均方根误差RMSE_tra＝0.457，训练集平均绝对误差MAE_tra＝0.329，误差值越小说明拟合程度越高，

说明模型具有较好的拟合优度和稳健性；外部验证采用预测与实测之间的拟合系数

和一致性相关系数CCC表示模型外部预测能力。判定依据：R²>0.7，Q²>0.6，R²-Q²<0.3，CCC>0.85；在本实施例中，最终的模型为

CCC＝0.930，表明模型具有很好的外部预测能力，图2给出模型的拟合程度及验证结果；The fit of the model is determined by the coefficient of determination corrected for the degrees of freedom

The training set root mean square error RMSE _tra and the training set mean absolute error MAE _tra characterize the coefficient of determination of the degree of freedom correction in this example

The training set root mean square error RMSE _tra = 0.457, the training set mean absolute error MAE _tra = 0.329, the smaller the error value, the higher the fitting degree,

It shows that the model has good goodness of fit and robustness; the external verification adopts the fitting coefficient between the prediction and the actual measurement

and the consistency correlation coefficient CCC represents the model's external predictive ability. Judgment basis: R ² >0.7, Q ² >0.6, R ² -Q ² <0.3, CCC>0.85; in this embodiment, the final model is

CCC=0.930, indicating that the model has good external prediction ability. Figure 2 shows the fitting degree and verification results of the model;

Step 5: The application domain of the model is characterized by the Williams diagram of the leverage value _hi based on the standard residual δ, specifically: when the absolute value of δ is generally considered to be greater than 3.0, the compound is an outlier, and when the leverage value _hi When it is greater than the warning value h*, it indicates that the structure of this compound is significantly different from other compounds; h _i and h* are calculated by the following formulas:

hi = x _i ^T (X ^T X) ^-1 x _i

h*=3(p+1)/n

where xi is the descriptor matrix of the ith compound; x _i ^T is the transpose matrix of x _i ; X is the descriptor matrix of all compounds; X ^T is the transpose matrix of X; (X ^T X) ^-1 is the matrix The inverse of X ^T X; p is the number of variables in the model and n is the number of data points in the dataset. As shown in Figure 3, the h* of the model is 0.054, and it is found that the model is suitable for predicting the value of logK _{PDMS-w for} compounds with hi less than 0.054.

Step 6 Model application: Use the model to predict the PDMS membrane/water partition coefficient of the unknown compound.

Example 1: Given a compound 3-chlorophenol, predict its log K _PDMS-w value. Firstly, the molecular structure of 3-chlorophenol was optimized according to the method of MM2 molecular mechanics. Secondly, based on the optimized molecular structure, the 4 molecular descriptors BLTD48, Hy, SpMaxA_B(s) and SpMaxA_AEA(dm) were calculated by alvaDesc 1.0.0 software. values, -3.341, -0.039, 0.789, and 0.374, respectively. According to the calculation formula, the hi value of the substance is 0.0353<0.054, so the compound is in the model application domain. Bringing the values of the above descriptors into the built model, the predicted value of log K _PDMS-w is 0.22, the experimental value is 0.31, and the predicted value is similar to the experimental value.

Example 2: Given a compound pentachlorophenol, predict its log K _PDMS-w value. Firstly, the molecular structure of pentachlorophenol was optimized according to the method of MM2 molecular mechanics, and then the values of its four molecular descriptors BLTD48, Hy, SpMaxA_B(s) and SpMaxA_AEA(dm) were calculated based on the optimized molecular structure using alvaDesc 1.0.0 software. , were -5.033, 0.078, 0.526 and 0.313, respectively. According to the calculation formula, the hi value of the substance is 0.0178<0.054, so the compound is in the model application domain. Bringing the values of the above descriptors into the established model, the predicted value of log K _PDMS-w is 2.68, the experimental value is 2.65, and the predicted value is similar to the experimental value.

The present invention is not limited to the above-mentioned embodiments. On the basis of the technical solutions disclosed in the present invention, those skilled in the art can make some substitutions and modifications to some of the technical features according to the disclosed technical contents without creative work. Modifications, replacements and modifications are all within the protection scope of the present invention.

Claims (8)

1. A method for predicting the membrane-water distribution coefficient of organic PDMS based on a quantitative structure-activity relationship model of SW-SVM algorithm is characterized by comprising the following steps:

step 1) data collection: the log K containing several organic compounds was collected from a review of the literature_PDMS-wValue, the resulting data set is log K_PDMS-wThe size of the value is extracted 1/5 as verification set data, and the rest is training set data;

step 2) descriptor computation: optimizing an initial molecular structure of the organic compound by using an MM2 molecular mechanics method, acquiring a molecular structure descriptor of the organic compound by using alvaDesc 1.0.0, and screening out a final descriptor through stepwise linear regression after pretreatment;

step 3), model construction: logarithm of organic PDMS-water distribution coefficient logK with final descriptor as independent variable_PDMS-wAdopting a support vector machine regression algorithm to establish a QSPR prediction model for the training set as a dependent variable, selecting optimized parameters through a k-fold cross validation algorithm, and establishing the QSPR model based on the optimal SW-SVM algorithm;

step 4), model verification: the model is verified, and the method comprises the following two steps: a) evaluating the goodness of fit and the robustness of the model; b) carrying out application domain representation and performance evaluation on the model; entering step 5) after the verification is qualified;

step 5) application domain characterization: characterizing the model application domain by a Williams diagram;

step 6) model application: the model was used to predict the PDMS film/water distribution coefficient of unknown compounds.

2. The prediction method according to claim 1, wherein in the step 1), the organic compound comprises polycyclic aromatic hydrocarbon, polychlorinated biphenyl, benzene, pesticide, ether, dioxin, ester, aliphatic, hydrocarbon, nitrogen sulfur compound.

3. The prediction method according to claim 1, wherein in step 1), data significantly deviating from the overall value of the collected same substance in the organic compound are removed, and the average value is used for model construction research.

4. The prediction method according to claim 1, wherein the organic compounds in the training set in step 1) are used for constructing a model, performing internal verification, and the organic compounds in the verification set are used for external verification of the model.

5. The prediction method according to claim 1, wherein in step 2) the preprocessing procedure comprises removing descriptors with constants, near constants, missing and correlation greater than 0.95.

6. The prediction method according to claim 1, wherein in the step 3), the QSPR model based on the optimal SW-SVM algorithm is constructed by using an R language package, and the method specifically comprises the following processes:

step 3-1, firstly dividing the whole data set into k sets, taking each set as a test set in turn, taking the rest sets as training sets, and repeating the training and testing for k times to ensure that each set is verified once as a test set;

step 3-2, calculating and comparing the average cross validation accuracy of k times of training, selecting a group of parameters with the highest cross validation accuracy, applying the group of parameters (cost, gamma) as the optimal value of k-fold cross validation to regression prediction of a support vector machine, wherein a penalty factor cost controls the relative proportion of model structure risk and experience risk, determines the superiority of the model, the gamma parameter determines the distribution of data after mapping to a new feature space, and the prediction model selects gamma as a radial basis kernel functionFormula g is 1/2 sigma²Wherein sigma is a width parameter of the function, and controls the radial action range of the function;

and 3-3, applying the parameters to the model to construct an optimized model.

7. The prediction method according to claim 1, wherein in the step 4, the goodness-of-fit and robustness evaluation indexes during model verification are as follows: coefficient of determination of degree of freedom correction

Training set root mean square error RMSE_traAnd training set mean absolute error MAE_tra。

8. The prediction method according to claim 1, wherein the step 5) specifically comprises: using a standard residual error based leverage value h_iThe Williams diagram of (1) characterizes the application domain of the model, with absolute values greater than 3.0, the compound being an outlier, with a lever value of h_iWhen the value is more than the alarm value h, the structure of the compound is obviously different from the structures of other compounds; h is_iAnd h is calculated by the following formula:

hi＝x_i ^T(X^TX)^-1x_i

h*＝3(p+1)/n

wherein x_iIs the descriptor matrix for the ith compound; x is the number of_i ^TIs x_iThe transposed matrix of (2); x is a descriptor matrix for all compounds; x^TIs the transpose of X; (X)^TX)^-1Is a matrix X^TThe inverse of X; p is the number of variables in the model and n is the number of data points in the dataset.

Images