Exploration of quantitative structure-reactivity relationships for the estimation of Mayr nucleophilicity

Quantitative structure-reactivity relationships (QSRRs) were investigated for the estimation of the Mayr nucleophilicity parameter N using data sets with 218 nucleophiles (solvent: CH₂Cl₂) and 88 compounds (solvent: MeCN) extracted from the Mayr's Database of Reactivity Parameters. The best predictions were observed for consensus models of random forests and associative neural networks, trained with empirical 2D and 3D CDK molecular descriptors, which yielded RMSE of 1.54 and 1.97 for independent test sets of the two solvent data sets, respectively. Compounds with silicon atoms were more difficult to predict, as well as classes of compounds with a reduced number of examples in the training set. The models' predictions were consistently more accurate than estimations simply based on the average of the N parameter within the class of the query compound. The possibility of calculating rate constants using the obtained models was also explored.