Aug-MIA-SPR/PLS-DA classification of carbonyl herbicides according to levels of soil sorption

Abstract

A major challenge in the design of new herbicides lies in the development of highly active, environmentally friendly compounds. Soil sorption is an ecotoxicological parameter used to probe the prospective environmental fate of persistent organic pollutants, such as some herbicides. This parameter, described in terms of logKOC (the logarithm of the soil/water partition coefficient normalized to organic carbon), is usually estimated using the octanol/water partition coefficient (logP, easily calculated or determined experimentally). However, estimations obtained with the logP are not always accurate. Thus, this work reports the use of molecular descriptors derived from multivariate image analysis of carbonyl herbicides to achieve a predictive classification model based on the partial least squares-discriminant analysis (PLS-DA) method. This model yields 80% accuracy in calibration, 75% in leave-one-out cross-validation and 100% in external validation. In addition, the Y-randomization test reveals that the obtained model is stable from fortuitous correlation, since the accuracy in calibration after shuffling the classes block is only 0.5%. Chemical interpretation in terms of the structural features that affect soil sorption is performed, based on the weights of the selected variables in the classification model. Finally, novel herbicides are rationally designed, based on the inferences arrived at in the structural interpretation experiment and predictions of their qualitative and quantitative soil sorption profiles performed, using the built aug-MIA-SPR and Wang's models, respectively.