Chemical applications of neural networks: aromaticity of pyrimidine derivatives.

Neural networks are computational tools able to apprehend non-linear relationships between different parameters, having the capacity to order a large amount of input data and transform them into a graphical pattern of output data. We have previously reported their use for the quantification of the aromaticity through the Euclidean distance between neurons. In this article, we apply the method to a variety of pyrimidine derivatives with electron-donor and electron-withdrawing groups as substituents, with capacity to produce push-pull compounds. We have calculated the aromaticity of benzene (as a reference molecule), parent pyrimidine and other 11 pyrimidine derivatives having amino, dimethylamino and tricyanovinyl substitution. The neural network has been generated using ASE, Λ, NICS(zz)(1) and HOMA as aromaticity descriptors, since our previous work showed that the combination of these indices provided the best performance of the network. On studying the influence of the substituent on the aromaticity of the molecule, we have found that, opposite to benzene derivatives, all the substituents decrease the aromaticity of the ring. The interplay between aromaticity, planarity and push-pull properties of all the substituted pyrimidines has also been addressed. An interesting feature of the neural network to quantify aromaticity is that the importance of the reference reaction used to evaluate energy stabilization and magnetic susceptibility exaltation is minimized.