Quantitative Structure-Activity Relationship Study on Thiosemicarbazone Derivatives as Antitubercular agents Using Artificial Neural Network and Multiple Linear Regression

Abstract

Background and purpose: Nonlinear analysis methods for quantitative structure–activity relationship (QSAR) studies better describe molecular behaviors, than linear analysis. Artificial neural networks are mathematical models and algorithms which imitate the information process and learning of human brain. Some S-alkyl derivatives of thiosemicarbazone are shown to be beneficial in prevention and treatment of mycobacterial infections and this study seeks to find out the relationship between structural features and the anti-tuberculosis activity of these compounds. Materials and methods: Multiple linear regression and Bayesian regularized artificial neural network (BRANN) for 47 compounds of thiosemicarbazone derivatives were designed using QSAR approaches. Descriptors were selected from a pool of 343 descriptors by stepwise selection and backward elimination. A three layer Bayesian regularized back-propagation feed-forward network was designed, optimized, and evaluated using MATLAB version R2009a. Results: The best model with 6 descriptors was found using multiple linear regression analysis: Log MIC= 2.592 + (0.067 ± 0.018) PMIX – (0.066 ± 0.017) PMIZ – (1.706 ± 1.600) Qneg – (0.235 ± 0.039) RDF030p + (0.118 ± 0.026) RDF 140u – (0.064 ± 0.021) RDF060p. The best BRANN model was a three-layer network with three nodes in its hidden layer. Conclusion: The BRANN model has a better predictive power than linear models and may better predict the anti-tuberculosis activity of new compounds with similar backbone of thiosemicarbazone moiety.

Keywords

• quantitative structure–activity relationship
• multiple linear regression
• bayesian regularized artificial neural network
• thiosemicarbazone derivatives