Studies of imidazole and pyrazole protonation using electrostatically trained neural networks

Abstract

A backpropagation neural network was trained with the molecular electrostatic potentials (MEPs) of a series of substituted imidazoles to predict their corresponding pK_a. Using MEPs determined with a variety of semiempirical and ab initio methods, the predictive power of the trained network was found to be sensitive to the quality of the basis set. The network was also trained to predict the proton affinity (E_pa) and pK_a, both individually and combined, for a series of pyrazole MNDO MEPs.