Abstract

Quantitative Structure-Retention Relationship models were developed to identify phenolic compounds using a typical LC- system, with both UV and MS detection. A new chromatographic method was developed for the separation of fifty-two standard phenolic compounds. Over 5000 descriptors for each standard were calculated using AlvaDesc software and then selected through Genetic Algorithm. The selected descriptors were used as variables for models construction and to obtain a better understanding of the retention behaviour of phenols during reverse-phase separation. Three distinct molecule sets, including fifty-two phenolic compounds (Set 1), 32 flavonoids (Set 2) and 15 mono-substituted flavonoids were divided into training and validation sets to build Partial Least Square, Multiple Linear Regression and Partial Least Square-Artificial Neural Network models. To assess the predictivity of the models, these were tested on a bergamot juice sample. Partial Least Square and Partial Least Square-Artificial Neural Network exhibit the lowest prediction error, and the latter showed the best predictive power in real sample recognition. The building and implementation of such predictive models showed to be a powerful tool to identify phenolic compounds based on retention data and avoiding the use of expensive and sophisticated detectors such as tandem MS.