Retention prediction of peptides based on uninformative variable elimination by partial least squares

A quantitative structure-retention relationship analysis was performed on the chromatographic retention data of 90 peptides, measured by gradient elution reversed-phase liquid chromatography, and a large set of molecular descriptors computed for each peptide. Such approach may be useful in proteomics research in order to improve the correct identification of peptides. A principal component analysis on the set of 1726 molecular descriptors reveals a high information overlap in the descriptor space. Since variable selection is advisable, the retention of the peptides is modeled with uninformative variable elimination partial least squares, besides classic partial least squares regression. The Kennard and Stone algorithm was used to select a calibration set (63 peptides) from the available samples. This set was used to build the quantitative structure-retention relationship models. The remaining 27 peptides were used as independent external test set to evaluate the predictive power of the constructed models. The UVE-PLS model consists of 5 components only (compared to 7 components in the best PLS model), and has the best predictive properties, i.e., the average error on the retention time is less than 30 s. When compared also to stepwise regression and an empirical model, the obtained UVE-PLS model leads to better and much better predictions, respectively.