Journal of computational chemistry

Chemical reactivity predictions: use of data mining techniques for analyzing regioselective azidolysis of epoxides.

PMID 20740561


Azidolysis of epoxides followed by reduction of the intermediate azido alcohols constitutes a valuable synthetic tool for the construction of beta-amino alcohols, an important chemical functionality occurring in many biologically active compounds of natural origin. However, depending on conditions under which the azidolysis is carried out, two regioisomeric products can be formed, as a consequence of the nucleophilic attack on both the oxirane carbon atoms. In this work, predictive models for quantitative structure-reactivity relationships were developed by means of multiple linear regression, k-nearest neighbor, locally weighted regression, and Gaussian Process regression algorithms. The specific nature of the problem at hand required the creation of appropriate new descriptors, able to properly reflect the most relevant features of molecular moieties directly involved in the opening process. The models so obtained are able to predict the regioselectivity of the azidolysis of epoxides promoted by sodium azide, in the presence of lithium perchlorate, on the basis of steric hindrance, and charge distribution of the substituents directly attached to the oxirane ring.