Phenothiazine and its derivatives are potent anticancer agents, these compounds inhibit cancer cells proliferation and tumor growth. A study of quantitative structure-activity relationship (QSAR) is applied to a set of 18 molecules derived from phenothiazine, in order to predict the anticancer biological activity of the test compounds and find a correlation between the different physic-chemical parameters (descriptors) of these compounds and its biological activity, using principal components analysis(PCA), multiple linear regression (MLR), multiple non-linear regression (MNLR) and the artificial neural network (ANN). We accordingly propose a quantitative model (nonlinear and linear QSAR models), and we interpret the activity of the compounds relying on the multivariate statistical analysis. Density functional theory (DFT) with Becke’s three parameter hybrid functional using the LYP correlation functional (B3LYP/6–31G (d)) calculations have been carried out in order to get insights into the structure, chemical reactivity and property information for the study compounds. The topological descriptors and the electronic descriptors were computed, respectively, with (ACD/ChemSketch; ChemBioOffice 14.0) and Gaussian 03W programs. A good correlation was found between the experimental activity and those obtained by MLR and MNLR respectively such as (R = 0,94 and R2 = 0,885) and (R = 0,986 and R2 = 0,973), this result could be improved with ANN such as (R = 0,988 and R2 = 0,976) with an architecture ANN (6-1-1). To test the performance of the neural network and the validity of our choice of descriptors selected by MLR and trained by MNLR and ANN, we used cross-validation method (CV) such as (R = 0,975 and R2 = 0,95) with the procedure leave-one-out (LOO). This study show that the MLR and MNLR have served to predict activities, but when compared with the results given by an 6-1-1 ANN model we realized that the predictions fulfilled by this latter was more effective and much better than other models. The statistical results indicate that this model is statistically significant and shows very good stability towards data variation in leave-one-out (LOO) cross validation.