Study Of Quantitative Structure-activity Relationship Of Nitrosamines Based On Artificial Neural Network And Support Vector Machine

N-Nitrosamines?NAs?are an important kind of harmful pollutants,which exhibit general carcinogenicity and teratogenicity to experimental animals.Exposure risk of NAs is inevitable owing to the wide existence of NAs in the environment intimately contacting with human,such as air,water and soil,as well as in foods and medicines.A large number of animal experiments and epidemiological studies showed that direct exposure to high doses of NAs could cause acute toxicity,and long-term exposure to low levels of NAs might induce malignant tumors.Therefore,the establishment of the quantitative structure-activity relationship?QSAR?between the molecular structure and physiological activity of NAs will be of great value for predicting the acute toxicity and carcinogenicity of NAs.In this study,the parameters of molecular structure acquired by quantum chemical method were employed as the descriptors and were screened by multiple linear regression and correlation analyses.Artificial neural network?ANN?and support vector machine?SVM?methods were employed for the construction of the QSAR models.The acute toxicity of 60 noncyclic NAs toward rats after oral administration were collected from Toxnet data base and used as the experimental data.Totally 9parameters were screened out as the descriptors,including polar molecular surface area?TPSA?,frontier orbital energy gap(E_LUMO-HOMO),octanol/water partition coefficient?log P?,N1-C4 bond length,N1-C5 bond length,N1-N2-O3 bond angle,molecular polarizability?IP?,Mulliken charge of N1?Q₁?and ATP charge of C4?ATP₄?,through sismultiple linear regression and correlation analy.A QSAR model with the structure of 9/10/1?input layer/hidden layer/output layer?was established using ANN method.The QSAR model was constructed with the data from 45compounds as the training sets,and then an internal validation of the model was performed.The results indicated that the model has satisfying internal predictability and robustness(R²_train=0.9561,Q²_LOO=0.9514).Subsequently,an external validation was performed for the model using 15 compounds as the test sets,and good external predictability was observed by achieving R²_test=0.9053 and Q²_ext=0.8842.The root mean square errors?RMSE?were obtained as 0.1534 and 0.2948 for the training and the test sets,respectively,which indicated that the model had good prediction accuracy.According to the QSAR model obtained from the acute toxicity of NAs,it was concluded that the acute toxicity of NAs was enhanced when halogen,benzyl group,or unsaturated bond in?-position existed within the side chain of NAs;while the acute toxicity was reduced when hydroxyl,carboxyl or phenyl group,or unsaturated bond in?-position existed within the side chain of NAs.Based on the above method for the QSAR model construction,the relationship between the molecular structure and the carcinogenicity of 38 noncyclic NAs was investigated.By sismultiple linear regression and correlation analysis,totally 6parameters including TPSA,E_LUMO-HOMO,log P,N1-N2-O3-C4 dihedral angle,total electronic potential?HF?and molecular volume?V?were screened out as the descriptors for the construction of a 6/6/1?input layer/hidden layer/output layer?QSAR model using ANN method.The results of the internal validation from the training set?30 compounds?were R²_train=0.7617 and Q²_LOO=0.7581,and the results of the external validation from the test set?8 compounds?were R²_test=0.7254 and Q²_ext=0.7088,which indicated that the model had favorable internal and external predictability and acceptable robustness.In the case of carcinogenicity of NAs,the shorter the side chain,the stronger the carcinogenicity was observed,which might be attributed to the low energy barriers for the metabolic activation of NAs with short chain;reduced carcinogenicity of NAs was observed when phenyl or polar groups were attached at the?-position of the side chain.In order to explore the influence of modeling method on the rationality of model,we further constructed the QSAR models using the SVM method for the two sets of data mentioned above.For the modeling of the acute toxicity of NAs,the QSAR model obtained by SVM method had favorable internal predictability,robustness and goodness of fit(Q²_LOO=0.9524 and R²_train=0.9580 from internal validation),which suggested that the model obtained via SVM method was comparable to that via ANN method.The prediction accuracy of the SVM models(RMSE_train=0.05358,RMSE_ext=0.2631)was better than that of the ANN model,while the external predictability of SVM model(R²_test=0.5130 and Q²_ext=0.5010)was worse than that of the ANN model.For the modeling of the carcinogenicity of NAs,the internal predictability(Q²_LOO=0.9440,R²_train=0.9450),external predictability(R²_test=0.9490,Q²_ext=0.9242)and the RMSE values(RMSE_train=0.033,RMSE_ext=0.2946)of the SVM model were all better than those of the ANN model,which indicated that the SVM method was more favorable than the ANN method for the modeling of small sample data.In summary,ANN and SVM methods were employed for the QSAR studies on the acute toxicity and carcinogenicity of noncyclic NAs.The obtained models were proved to have satisfying predictability,accuracy,goodness of fit and robustness by performing internal and external validation.This study not only provided reliable methods for predicting the acute toxicity and the carcinogenicity of NAs,but also has important significance for revealing the the toxicity/carcinogenesis of NAs and for reducing the environmental health risks of NAs.