Classification of gas chromatographic fingerprints of saffron using partial least squares discriminant analysis together with different variable selection methods

In the present work, the abilities of five different variable selection methods including recursive partial least squares (rPLS), variable importance in projection (VIP), selectivity ratio (SR), significance multivariate correlation (sMC), and PLS loading weights were evaluated on the supervised classification of gas chromatographic fingerprints of saffron using PLS-discriminant analysis (PLS-DA). In this regard, eighty-three saffron samples analyzed by gas chromatography-flam ionization detector (GC-FID), were used as a case study. The GC-FID chromatograms of saffron samples were baseline corrected and aligned using asymmetric least squares (AsLS) and correlation optimized warping (COW) methods, respectively. Then, the whole digital profiles of preprocessed chromatograms were normalized to internal standard (I.S.), mean-centered, pareto-scaled and finally modeled by PLS-DA to classify saffron samples according to their cultivation areas. Afterwards, performance of different variable selection methods (i.e., rPLS, VIP, SR, sMC and loading weights) for choosing the most important variables (i.e., retention time points) in GC-FID fingerprints, were compared in terms of the model's interpretability and predictability. The results indicated that although different variable selection methods could select different subset of variables, but, prediction ability of all the models were still acceptable. The best model performance was achieved when the result of all variable selection methods were taken into account. Finally, nine secondary metabolites of saffron suggested by almost all selection methods were chosen as the saffron biomarkers.