Prediction of a biological variable (Ames test on TPM) and of a chemical variable (TSNAs in smoke) according to the tobacco's chemical and physical characteristics
Chemical analyses of special analytes (e.g. TSNAs) or biological tests on the smoke are rather sophisticated and time consuming technologies. Giving estimates based on fast analyses would enable the blenders to have a tool allowing them to reduce tobacco health risks. Among the biological tests, we have taken as variable to predict a test of TPM mutagenicity (Ames test on Salmonella typhimurium, strain TA98 with S9 activation) and as chemical variable, TSNAs in mainstream smoke. The analyses have been carried out on several types of tobaccos (flue-cured, Oriental, burley, dark) from various geographical origins. A big number of chemical, physical and spectral characteristics (NIRS) have been measured as predictive variables. Constant NTM cigarettes have been made and smoked in order to make the Ames tests as well as to determine the TSNAs in the mainstream smoke. Several predictions methods have been applied: Multiple Linear Regression, Partial Least Square regression (PLS) per bloc of variables, PLS multiblocs, in order to conceive models and to test them on new individuals. Different models will be presented and their validity will be analysed and discussed according to the values of Residual Mean Square Error of Estimation and Residual Mean Square Error of Prediction taking into account the reproducibility of the variable to predict.