QSAR Models for Predicting Additive and Synergistic Toxicities of Binary Pesticide Mixtures on Scenedesmus Obliquus①

MO Ling-Yun YUAN Bi-Kng ZHU Jie QIN Li-Tng,b② DAI Jun-Feng,d②

a (College of Environmental Science and Engineering, Guilin University of Technology, Guilin 541004, China)

b (Guangxi Key Laboratory of Environmental Pollution Control Theory and Technology,Guilin University of Technology, Guilin, Guangxi 541004, China)

c (Technical Innovation Center of Mine Geological Environmental Restoration Engineering in Southern Karst Area, MNR, Nanning 530023, China)

d (Guangxi Collaborative Innovation Center for Water Pollution Control and Water Safety in Karst Area, Guilin University of Technology, Guilin 541004, China)

ABSTRACT Pesticides released into the environment may pose potential risks to the ecological system and human health. However, existing toxicity data on pesticide mixtures still lack, especially regarding the toxic interactions of their mixtures. This study aimed to determine the toxic interactions of binary mixtures of pesticides on Scenedesmus Obliquus (S. obliquus) and to build quantitative structure-activity relationship models (QASR) for predicting the mixture toxicities. By applying direct equipartition ray method to design binary mixtures of five pesticides (linuron, dimethoate, dichlorvos, trichlorfon and metribuzin), the toxicity of a single pesticide and its mixture was tested by microplate toxicity analysis on S. obliquus. The QASR models were built for combined toxicity of binary mixtures of pesticides at the half-maximal effective concentration (EC50), 30% maximal effective concentration (EC30) and 10% maximal effective concentration (EC10). The results showed that the single toxicity follows:metribuzin ＞ linuron ＞ dichlorvos ＞ trichlorfon ＞ dimethoate. The mixtures of linuron and trichlorfon, dichlorvos and metribuzin, dimethoate and metribuzin induced synergetic effects, while the remaining binary mixtures exhibited additive. The developed QSAR models were internally validated using the leave-one-out cross-validation(LOO), leave-many-out cross-validation (LMO), bootstrapping, and y-randomization test, and externally validated by the test sets. All three QSAR models satisfied well with the experimental values for all mixture toxicities, and presented high internally (R2 and Q2 ＞ 0.85) and externally (Q2F1, Q2F2, and Q2F3 ＞ 0.80) predictive powers. The developed QSAR models could accurately predict the toxicity values of EC50, EC30 and EC10 and were superior to the concentration addition model's results (CA). Compared to the additive effect, the QSAR model could more accurately predict the binary mixture toxicities of pesticides with synergistic effects.

Keywords: pesticide, QSAR, toxicity prediction, binary mixture, algae;

1 INTRODUCTION

In recent years, with the improvement of agricultural science and technology and the increase of agricultural inputs,pesticides play a great role in agrarian cultivation, such as improving crop yield[1]. The threats posed to the environmental quality of agricultural soils and surface water cannot be ignored. Studies have shown that nonpoint source pollution is the leading cause of surface water pollution globally,with the largest contribution from agricultural point source pollution. Pesticide loss and residual pollution caused by the heavy application of pesticides, which causes the environmental pollution of soil and water in agricultural fields, have become an increasingly prominent ecological problem[2,3].For example, Zheng et al.[4]detected 82 pesticide residues in the Chiu-lung Rive of Fujian Province, China. Canccapa et al.[5]found high detection rates of chlorpyrifos, diazinon, and carbendazim in the Turia and JúcarRivers (Spain) from 2010 to 2013, respectively. Xu et al.[6]reported the results of pesticide residue monitoring from 55 main water sources in 12 urban areas of Yantai City: organophosphorus and pyrethroid pesticides were detected in all water sources at concentrations ranging from 103.0 to 345.7 ng/L. Therefore, it is of great significance to study the pollution and ecological risk of pesticide exposure to the environment.

Algae play an essential role in aquatic and soil ecosystems.Algae can carry out photosynthesis, and the characteristics of their species can directly affect the structures and functions of aquatic ecosystems, which has been used as indicator organisms in ecotoxicological research in recent years[7]. Many research achievements have been obtained. Tien et al.[8]studied the single toxic effects of chlorpyrifos, terbufos and methamidophos on diatoms, cyanobacteria and green algae.The results showed that different algae showed wide variation in sensitivity to different pesticides, with green algae being the most tolerant withEC50of 1.29～41.16 mg/L. Wan et al.[9]investigated the toxicity of trichlorfon to the freshwater algaChlamydomonas reinhardtii, with anEC50of 200 mg/L. Studies have shown thatChlamydomonas reinhardtiihas a high tolerance ability to trichlorfon and is promising for removing trichlorfon from natural water environments. Liu et al.[10]studied the single and combined toxicities of six pesticides (simetryn, bromacil, hexazinone, dodine, propoxur, and metalaxyl) onChlorella pyrenoidosa. With the concentration addition as an additive reference model, four kinds of binary mixture rays exhibited antagonistic effects onChlorella pyrenoidosa. It is imperative to apply algae to the toxicity research of pesticide pollutants, and it is of great significance to evaluate the combined toxicity of pesticide pollutants for ecological risk.

There are many varieties of pesticides, and compound pesticides account for a large proportion. Mixed-use, blind use or abuse of multiple pesticides lead to the prominent phenomenon of multiple pesticide residues in the environment. This leads to the fact that pesticides in the environment will mix in many and varied forms, resulting in complex mixture systems[11]. It is possible that the components of these complex mixtures produce toxic interactions with each other or are enriched in living organisms, causing more serious environmental pollution problems. Therefore, the investigation of combined toxic effects of pesticides has become the focus of researchers in recent years. Tien et al.[8]studied the combined toxicity of chlorpyrifos, terbufos and methamidophos on diatoms, cyanobacteria and green algae. The results indicated that the mixtures of pesticides exhibited antagonism and synergism to algae, and pesticide mixtures are more likely to induce detoxification mechanisms than single pesticides. Du et al.[12]used plasma metabolomics to evaluate the toxic effects of four organophosphorus pesticides (dichlorvos, acephate, dimethoate and phorate) on male Wistar rats. The results showed that the combination of the four drugs could produce a combined effect at a non-damaging effect level,causing body oxidative stress and liver and kidney dysfunction.

Quantitative structure-activity relationship (QSAR) is a method that can effectively predict the physicochemical properties, environmental behavior, and toxicity characteristic effect parameters of organic pollution[13]. The qualitative identification of the combined effect of mixtures has achieved remarkable results. Qualitative research can determine that the combined effect of mixtures is antagonism,synergism or additive[14,15], which has been widely used.However, this qualitative discrimination cannot quantitatively determine the intensity of toxicity, so the research on combined toxicity must be changed from qualitative research to quantitative research. QSAR is one of the commonly used methods to study the combined toxicity of mixtures quantitatively, and it is also the frontier and emphasis of current structural-activity relationship research[16]. In this study, we selected the green algaeScenedesmus obliquus(S. obliquus)as target organisms and determined the toxicities of five pesticides, including two herbicides (linuron and metribuzin) and three insecticides (dimethoate, dichlorvos and trichlorfon).Based on the single toxicity of five pesticides and the combined toxicity of seven binary mixture systems designed by direct equipartition ray (EquRay)[17], as well as screening the optimal descriptors describing the contributions of these five pesticides to the combined toxic effect, QSAR models that could accurately predict the toxic value of binary mixtures were developed. Therefore, the model may serve as a theoretical basis for predicting the binary toxicities of pesticide compounds.

2 MATERIALS AND METHODS

2. 1 Test organism and culture

The insecticide and herbicide are used in agricultural areas to protect crops and control pests and weeds. Linuron and metribuzin belong to herbicides, and dimethoate, dichlorvos and trichlorfon belong to insecticides, and are widely used in agricultural production[18]. However, in practical use, their actual utilization is very low, and a large amount of the residual herbicides and insecticides enter into the ecosystem such as soil and water bodies[19], which pose a serious threat to the ecological environment and even human health. Researchers have paid great attention to the pollution of herbicides and insecticides to the environment[20], but there are few reports on their mixtures' toxic effects. Five pesticides including linuron (CAS 330-55-2, purity ＞ 99.44%), dimethoate (CAS 60-51-5, purity ＞ 99.30%), dichlorvos (CAS 62-73-7, purity ＞ 99.62%), trichlorfon (CAS 52-68-6, purity ＞97.20%), and metribuzin (CAS 21087-64-9, purity ＞ 99.82%)were purchased from Dr Ehrenstorfer GmbH Co.. TheS.obliquuswas purchased from Freshwater Algae Culture Collection at the Institute of Hydrobiology (FACHB), numbered FACHB-5.

2. 2 Toxicity test and concentration-response curve fitting

A modified procedure of the microplate toxicity analysis(MTA) based on algae growth inhibition was used to determine the toxicity of a pesticide or a mixture toS. obliquus[21].In the MTA, the peripheral wells were filled with 200 μL water to minimize the edge effects. A total of 24 wells of the second, sixth, seventh, and eleventh columns in a 96-well microplate were selected as blank controls, and the remaining 36 microwells were designed with twelve concentration gradients of three parallels by a dilution factor. A total of 100 μL algal liquid was added to each well, and the total volume was 200 μL. Three plates were repeated and incubated at 25 ℃with a light:dark ratio of 12:12. The optical density (OD) ofS.obliquuswas determined on the Power Wave microplate spectrophotometer (American BIO-TEK Company) after 96 h at the wavelength of 690 nm. The toxicity is expressed in terms of inhibition rate (E), and the formula is given in Eq.(1).

whereODwas the average optical density of the experimental group andOD0was that of the control group.

To obtain the effect value and effect concentration for individual pesticide or their mixtures, Logit (Eq. 2) and Weibull(Eq. 3) functions were used to fit the concentration-response curve (CRC) measured by MTA with the nonlinear leastsquares method. In statistics, the ratio of the number of data points (i. e., the number of samples) to the model parameters should be greater than 5:1. There were 12 concentration points in the MTA. Therefore, the two-parameter equation(Logit and Weibull) was used to fit the curve in this study,which could effectively reduce the over-fitting phenomenon.The coefficient of determination (R2) is an evaluation index of the final fitting regression effect. For each CRC, Logit and Weibull functions were used to fit the curve simultaneously,and then theirR2was compared. The closerR2is to 1, the better the fitting regression effect is.

whereEis the toxic effect value,cthe concentration dose,αthe positional parameter, andβthe slope parameter.

2. 3 Mixture design

To systematically examine the variation of toxicity in mixtures, five mixtures ray with different concentration ratios (pi,i= 1, 2, 3, 4, 5) were designed by direct equipartition ray design (EquRay) for every group mixture (pi, Table 2), and 12 concentration points were arranged for each ray. The concentration ratio (pi) of a component (pesticide) is defined as a ratio of the concentration of the component in a mixture to the sum of concentrations of all components in the mixture.There are seven groups of binary mixtures, B1 consisting of linuron and dimethoate, B2 of linuron and dichlorvos, B3 of linuron and trichlorfon, B4 of dichlorvos and trichlorfon, B5 of dichlorvos and metribuzin, B6 of trichlorfon and metribuzin and B7 of dimethoate and metribuzin.

2. 4 Toxic interaction analysis of the mixtures

The concentration addition (CA) model was used as the additive reference model to analyze and compare the toxic interactions in different concentration ranges from CRCs of mixtures. The formula of the CA model is expressed as in the following equation[22]:

whereECxmixis the mixture concentration inx% combined effect,nis the number of mixture components,ECx,irepresents the concentration at which the effect (x%) produced in the presence of theithcompound alone in the mixture, andpiis the concentration ratio of theithcomponent in the mixture.

The combined toxic interaction of mixture was qualitatively distinguished by comparing experimental CRC with predictive CRC by CA in the whole effect range[23]. Due to the experimental error in the toxicity experiment and the fitting error in the nonlinear fitting of CRC, 95 percent observed confidence intervals (OCI) of CRC must be considered to compare the predictive CRC by CA with the experimental CRC. The predictive CRC by CA is almost located between the upper limit and lower one of the 95 percent OCI, which states that the toxicities of the mixtures are additive. The predictive CRC by CA significantly deviates from the experimental CRC, locating above the 95 percent OCI and exhibiting antagonistic interaction. The predictive CRC by CA significantly deviates from the experimental CRC, locating under the 95 percent OCI and

displaying synergistic interaction.

2. 5 Characterization of molecular structure of mixtures

Firstly, 3Dmolecular structures of five single pesticides and their binary pesticide mixtures were pre-optimized for energy minimization of Guassian09 software[24]. Secondly,density functional theory (DFT) of Guassian09 software at the B3LPY/6-31G (d,p) level was employed to optimize the molecular structure of single and binary mixed pesticides to the best transition state[24]. Sixteen descriptors were extracted,including polarizability, single-point energy, dipole moment,the most positive charge, the most negative charge, EHOMO,ELUMO, zero vibration energy, enthalpy of formation, Gibbs free energy, heat correction value, constant volume molar melting, entropy, absolute hardness, softness, and chemical potential. Thirdly, the AM1 method which runs the MOPAC program was used to calculate eight descriptors of molecular structures of single and binary mixed pesticides, including the final formation heat, total energy, electron energy, core repulsion, COSMO surface area, COSMO volume, ionization potential and molecular weight[25]. Finally, Dragon software[26]can only calculate the molecular descriptors of a single pesticide, including topological index, connection index, RDF descriptor, 3D-MoRSE descriptor, and edge adjacency index. A total of 5270 single pesticide molecular structure descriptors are screened preliminarily in dragon software, and the remaining 43 qualified descriptors are obtained.

If the binary mixture conforms to the concentration addition model, it can be considered that the mixture has no interaction, which can be characterized by the mixture descriptor (xmix) as follows[27]:

wherexmixis mixture descriptors,xirepresents a descriptor for component, andpiis the concentration ratio of theithcomponent in the mixture.

If the components contained in the mixture are considered together as a whole, and molecular structure descriptors of all components in the mixture are calculated simultaneously by Guassian09 and MOPAC software, a mixture descriptor with interaction information can be obtained as below:

wherexmixis the mixture descriptors,xirepresents a descriptor for component,piis the concentration ratio of theithcomponent in the mixture,nis the number of components of the mixture andxmis the descriptor obtained when the mixture is calculated as a whole.

2. 6 Establishment and validation of the QSAR model

The overall data set is randomly divided into a training set(70%) and a test set (30%) in QSARINS software[28,29]. The test set is used for external validation of the model. The best descriptor based on genetic algorithms was selected in QSARINS software and QSAR models were built with training set by multiple linear regressions (MLR). To ensure the reliability and validity of the regression model, the leave-one-out cross-validation (LOO), leave-many-out cross-validation(LMO),y-randomization test, and the bootstrapping method are used for the internal validation.

The application domain (AD) of the QSAR model is a thoretical area in the space defined by the descriptors used in the model, and established QSAR models should make predictions about the reliability of that space, defining the AD of the QSAR model with a leverage valuehi, which is calculated as[30]:

wherexirepresents the original variable of molecular descriptor for a compound andXis the descriptor matrix of the compounds of the training set.

3 RESULTS AND DISCUSSION

3. 1 Toxicity of five pesticides to S. obliquus

The fitted CRCs parameters (αandβ), fitted statistics (R2andRMSE), andEC(10,30,50)of five pesticides are listed in Table 1. From Table 1, apart from the CRCs of dimethoate and dichlorvos depicted by the Weibull function, the CRCs of the other three pesticides toS. obliquuscan be well described by the Logit function (R2＞ 0.9704 andRMSE＜ 0.0783). The experimental concentration-toxicity (inhibition percent) data points and fitted CRCs are shown in Fig. 1. TheEC50value is considered as a toxicity index, and the toxicity order of five pesticides toS. obliquuswas metribuzin ＞ linuron ＞ dichlorvos ＞ trichlorfon ＞ dimethoate. The freshwater green algaeS.obliquuswas more sensitive to two herbicides (linuron and metribuzin) than the insecticide. The structures of dimethoate,dichlorvos, and trichlorfon have a similar frame-work, except that the side chain substituents are different, and the toxicities of dichlorvos and trichlorfon containing chlorine substituents were slightly greater than that of dimethoate without chlorine substituents. Therefore, the side-chain chlorine substituents of pesticide molecules may enhance their toxicity. In contrast,metribuzin is close to and has a cyclic structure to linuron,and its toxicity is higher than that of dimethoate, dichlorvos,and trichlorfon, so the cyclic structure of pesticide molecules may increase their toxicity.

Table 1. Fitted CRC Parameters (α and β), Fitted Statistics (R2 and RMSE), and EC(10, 30, 50) of Five Pesticides

Fig. 1. Concentration-response curves of five pesticides to S. obliquus

3. 2 Combined toxicity of pesticide mixtures

Using the MAT method to determine the toxicity of 35 rays of pesticide binary mixture toS. obliquus, the experimental concentration-toxicity (inhibition percent) data points and fitted CRCs are shown in Fig. 2. The CRCs of two rays (B1-R1 and B3-R5) of 35 mixture rays can be well described by Logit (L) functions and the other 33 rays can be described by Weibull (W) functions (Table 1S). The fitted regression coefficients (αandβ) and statistics, determination coefficient (R2)and root mean squared error (RMSE) as well aspEC(50,30,10)are listed in Table 1S. It has been shown that 35 mixture rays exhibit a good statistical significance with the determination coefficient of ＞ 0.93 and the root mean squared errorof ＜0.1254, which explains that all pesticide mixtures have a good concentration-response relationship.

3. 3 Toxicity interaction between pesticide mixtures

Selecting CA as an additive reference model, if the experimental fitting effect is higher, equal to or lower than the CA prediction effect, the mixture exhibits synergistic, additive and antagonistic interactions. From Fig. 2, at different concentration proportions, the fitted CRCs deviated from CA predicted line to different degrees, i.e., the different interactions.

The predictive CRCs by CA of five binary mixture rays,B3-R1, B3-R2, B7-R2, B7-R3 and B7-R5, significantly deviated from the experimental CRCs, locating under the 95 OCI and displaying synergistic interactions. The predictive CRCs by CA of ten binary mixture rays, B3-R3, B3-R4, B3-R5,B5-R1, B5-R2, B5-R3, B5-R4, B5-R5, B7-R1, and B7-R4,evidently deviated from the experimental CRCs in high concentration area (includingEC50andEC30), locating under the 95 percent OCI and showing high concentration synergistic interactions. The predictive CRC by CA of B6-R1 clearly deviated from the experimental CRC in the medium concentration zone (includingEC10), locating under the 95 percent OCI and exhibiting medium concentration synergistic interactions. The predictive CRC by CA of B2-R1 obviously deviated from the experimental CRC in the low concentration region, locating under the 95 percent OCI and demonstrating low concentration synergistic interactions. For the remaining 18 rays of 35 mixture rays, the predictive-CRCs were almost completely located between the upper limit and lower one of the 95 percent OCI, which states that the toxicities of the mixtures are additive. Furthermore, 16, 15, and 10 rays of mixtures displayed synergism atEC50,EC30andEC10concentration effects, respectively.

The CA model was used to evaluate the toxicity interaction(synergism or additive) of binary mixtures of pesticides. The results showed that the proportions of rays with synergistic effect were 46%, 43% and 29% atEC50,EC30andEC10concentrations, respectively. Therefore, the variation of concentration level may change the toxicity interaction of binary mixtures. Besides, fromEC10toEC50, the synergistic effect of pesticides onS. obliquuswas enhanced with the increase of concentration level of a binary mixture. As previous studies have shown, synergy was enhanced at the high concentration level of binary mixtures such as pesticides and metals[31,32].

Fig. 2. Concentration-response curves of 35 rays of pesticide binary mixture to S. obliquus

3. 4 QSAR model of pesticide mixtures

For genetic algorithm parameters, we selected the maximum population size (1000 chromosomes), the minimum mutation rate (0.05), the maximum model size (4 variables)and 500 as the number of generations. By selecting four descriptors as the best independent variables and using thepEC50,pEC30andpEC10(the negative logarithm ofEC50,EC30andEC10) values as dependent variables, the QSAR models of pesticide mixtures were established, including modelpEC50, modelpEC30and modelpEC10. The modelpEC50was shown as follows:

where, TE is the total energy, COA the cosmo area, DM the dipole moment, GGI4 the topological charge index of order 4,RMSEthe root mean square error of the model,Fthe Fischer’s statistic,R2the coefficient of multiple determination,Q2LOOthe correlation coefficient of leave-one-out crossvalidation, andQ2LMOthe correlation coefficient of leavemany-out cross-validation,R2bstrandQ2bstrare the average correlation coefficient of bootstrapping method,R2yrandandQ2yrandthe maximumR2andQ2the values of they-randomization tests,RMSPextis the root mean square error of test set,Q2F1，Q2F2andQ2F3are the external validation coefficient of the model, andCCCis the concordance correlation coefficient of the model. The values before and after“±” represent the regression coefficient of the model and its corresponding standard deviation.

The modelpEC30equation is as follows:

where, COA is the cosmo area, MW the molecular weight,EHOMOthe highest molecular orbital energy, and GGI4 the topological charge index of order 4.

The statistic is used to evaluate the internal and external predictive ability of QSAR models. If the statistic values are satisfied (thus the models are robust), the models are acceptable.Q2greater than 0.5 considered the model to be good,and greater than 0.9 showed the model excellent[33]. Tropsha et al.[34]suggested that bothR2andQ2are greater than 0.6.There was no significant difference between correlation coefficients of modelspEC50(R2= 0.9403 andQ2LOO= 0.8321),pEC30(R2= 0.9298 andQ2LOO= 0.8780) andpEC10(R2=0.8564 andQ2LOO= 0.7615), which shows these models have strong robustness. In addition,R2-Q2LOO＜ 0.1[35], so the models have no over-fitting. Furthermore, the statistic values ofQ2LMOandQ2LOOwere close to each other, and the result of bootstrapping method[36]is as below:R2bstr＞ 0.6,Q2bstr＞0.5, indicating that the models all have good robustness. The result ofy-randomization test wasR2yrand＞Q2yrand[37], which demonstrates no chance correlation between the independent and dependent variables of the models. The external predictive ability of the model can be assessed with statistics ofQ2F1,Q2F2andQ2F3. All relevant statistical parameters of the external validation of the three models meet the conditions suggested by Chirico and Gramatica[38], in which allQ2F1,Q2F2, andQ2F3shall be greater than 0.6 and the concordance correlation coefficient (CCC) should be greater than 0.85. It could be indicated that the three models have a good external predictive ability.

Williams plot is the relationship between LOO standard residual and leverage value[39], and the leverage threshold (h*)of modelspEC50,pEC30andpEC10was 0.4800. The leverage thresholdh* = 0.4800 was derived according to QSARMLR program, and the standard residual is the value obtained by dividing the residual by its standard deviation, setting at ±2.5[40]. If the absolute value of the LOO standard residual of a sample is greater than 2.5 and greater thanh*, the sample is abnormal[41]. Moreover, it could be seen from Fig. 3 that the calculated values of the training set and the predicted values of the test set are all within the application domain, indicating that the model has no abnormal samples and the data of all binary mixture rays are reliable.

Fig. 3. Williams plot based on standardized residuals for three models

3. 5 Interpretation of the descriptors

GGI4 is defined as the sum of the corresponding charge terms absolute values over a pair of vertices (nonhydrogen atoms) with all topological distances equal to 4 in the molecule. Its value is mainly determined by the number of nonhydrogen atoms, the charge, and the topological distance. GGI4 can be calculated by establishing the compound matrix (distance matrixD, coulomb matrixT, matrixM=C⋅T, whereCis the adjacency matrix). The mathematical model can be expressed as[42]:

where elementdijof matrixDis the topological distance between atomsiandj,gij=mij-mji(mijis the matrixMelement) denotes the difference in magnitude of net charge of atomsitoj,Nis the number of nonhydrogen atoms in the molecule,kis an indicator variable for the topological distance between atomsiandj, taking values 1～10k(k= 4 in this descriptor),δis Kronecker delta[43], defined as two variables,iandj, which equals 1 wheni=jand equals 0 wheniis not equal toj.

The cosmo area is used in the cosmo model to characterize the size of the area where solvent molecules surround the molecules in solution[44]. Total energy includes electronic energy and nuclear repulsion energy, and its value is mainly related to the number of atoms in the molecule and the charge the atom itself carries. Dipole moment belongs to the geometric descriptor[45,46], a physical quantity that describes the situation of charge distribution in a molecule. It usually contains information about the property and location of all atoms and bonds in the molecule, representing properties such as geometry, type of chemical bonds and so on of the molecule.

Molecular weight is the sum of the relative atomic mass of all atoms in a molecule. EHOMOrepresents the molecular orbital with the highest energy occupied by electrons. With the increase of atomic number, the effective nuclear charge and the number of the main quantum also increases, leading to the electron moving towards the orbital of higher energy level, so that the molecular orbital energy is affected by the type of atoms, the number and the spatial distance of the atoms[47].

3. 6 Model comparison between QSAR and CA

The QSAR model is used to quantitatively analyze the toxicity of mixtures based on the molecular structure information of mixtures, while the CA model is used to qualitatively analyze the toxicity of mixtures based on the toxicity of single compounds[48]. Therefore, their toxicity prediction ability on binary mixtures also has a large gap. Moreover, it could be seen from Fig. 4 that the QSAR model has a higher fitting degree than the CA model. TheRMSEof QSAR model atEC50,EC30andEC10were 0.1033, 0.1196 and 0.1759, respectively, while 0.4273, 0.4182 and 0.4818 for the CA model. It could be concluded that the experimental value of the CA model has a great deviation from the predicted one, and the QSAR model has higher toxicity predictive ability than the CA model at all three concentration effects. In addition,modelspEC50andpEC30had higherR2(0.9403 and 0.9298)and lowerRMSPext(0.142 and 0.144), and their predictive ability and robustness were significantly better than those of modelpEC10(R2= 0.8564 andRMSPext= 0.1781). By combining section 3.1 analysis, 16, 15 and 10 binary mixture rays with synergistic effect could be seen in modelspEC50,pEC30andpEC10, respectively, which shows that with the increase of the synergistic effect of the mixture, the predictive ability of established QSAR model is more excellent. Thus, the QSAR model can predict the synergistic toxicity value of binary mixtures more accurately.

Fig. 4. Plot of the observed versus predicted pEC50, pEC30 and pEC10 resulted from QSAR model (A, B, C) and CA model (D, E, F)

4 CONCLUSION

The concentration-response curves ofS. obliquusfor five pesticides and their 35 binary mixture rays could all be well fitted by Logistic or Weibull functions. The toxicity order of five pesticides was metribuzin ＞ linuron ＞ dichlorvos ＞ trichlorfon ＞ dimethoate. The freshwater green algaeS. obliquuswas more sensitive to two herbicides (linuron and metribuzin)than the insecticide. The toxicity of a single pesticide may be related to its cyclic structure. Metribuzin and linuron with cyclic structure had higher toxicity than other pesticides.

The mixtures of linuron and trichlorfon, dichlorvos and metribuzin, dimethoate and metribuzin exhibited synergistic effects, while the remaining binary mixtures exhibited additive effects. Selecting CA as an additive reference model to evaluate the toxicity interaction of binary mixtures of pesticides could found that 46%, 43% and 29% of the binary mixtures have synergistic effects atEC50,EC30andEC10concentrations, respectively, indicating that the interaction model of chemicals may change with the variation of mixture concentration levels or concentration ratios, and the higher the concentration level, the greater the possibility of synergistic effect.

Mixed structure descriptors can effectively characterize the overall molecular structure of the mixture. The QSAR models established at the effect concentrations ofEC50,EC30andEC10had good internal robustness (bothR2andQ2＞ 0.8) and external predictive ability (Q2F＞ 0.8,CCC＞ 0.85). The excellent degree of the model was modelpEC50＞ modelpEC30＞ modelpEC10, and the predictive ability was better than that of the CA model under the same concentration effect. Therefore, three models may be useful in predicting unknown toxicity values of the binary pesticide mixtures. In addition,compared to additive, three models can more accurately predict the toxicity of binary mixtures of pesticides with a synergistic effect.