Artificial neural networks (ANN; feed-forward mode) are used to quantitatively estimate the enantioresolution (Rs) in cellulose tris(3,5-dimethylphenylcarbamate) of chiral molecules from their structural information. To the best of our knowledge, for the first time, a dataset of structurally unrelated compounds is modelled using ANN, attempting to approach a model of general applicability. After setting a strategy compatible with the data complexity and their relatively limited size (56 molecules), by prefixing initial ANN inner weights and the validation and cross-validation subsets, the ANN optimisation based on a novel quality indicator calculated from 9 ANN outputs allows selecting a proper (predictive) ANN architecture (a single hidden layer of 7 neurons) and performing a forward-stepwise feature selection process (8 variables are selected). Such relatively simple ANN offers reasonable good general performance in predicting Rs (e.g. validation plot statistics: mean squared error = 0.047 and R = 0.98 and 0.92, for all or just the validation molecules, respectively). Finally, a study of the relative importance of the selected variables, combining the estimation from two approaches, suggests that the surface tension (positive overall contribution to Rs) and the -NHR groups (negative overall contribution to Rs) are found to be the main variables explaining the enantioresolution in the current conditions.
Keywords: Artificial neural networks; Cellulose chiral stationary phase; Enantioresolution; Feature selection; Heterogeneous dataset; Relative variable importance.
Copyright © 2022. Published by Elsevier B.V.