Abstract

Combinations of multiple linear regressions, genetic algorithms and artificial neural networks were utilized to develop models for seeking quantitative structure-activity relationships that correlate structural descriptors and inhibition activity of adenosine deaminase competitive inhibitors. Many quantitative descriptors were generated to express the physicochemical properties of 70 compounds with optimized structures in aqueous solution. Multiple linear regressions were used to linearly select different subsets of descriptors and develop linear models for prediction of log(k(i)). The best subset then fed artificial neural networks to develop nonlinear predictors. A committee of six hybrid models - that included genetic algorithm routines together with neural networks - was also utilized to nonlinearly select most efficient subsets of descriptors in a cross-validation procedure for nonlinear log(k(i)) prediction. The best prediction model was found to be an 8-3-1 artificial neural network which was fed by the most frequently selected descriptors among these subsets. This prediction model resulted in train set root mean sum square error (RMSE) of 0.84 log(k(i)) and prediction set RMSE of 0.85 log(k(i)) (both equivalent of 0.10 in normal range of log(k(i))) and correlation coefficient (r(2)) of 0.91.