The calibration set consisted of a total of 116 samples (33 samples of roasted coffee, 27 samples of roasted coffee husks, 30 samples of roasted corn and 26 samples of adulterated coffee, with adulteration levels ranging from 50 to 10% of one or both adulterants). The evaluation set consisted of a total of 49 samples (15 samples

of roasted coffee, 11 samples of roasted coffee husks, 16 samples of roasted corn and 7 samples of adulterated coffee, with adulteration levels ranging from 50 to 10% of one or both adulterants). For both the calibration and evaluation sets, each sample represented one spectra, without any averaging procedure. It was observed that model recognition ability varied significantly with the number of variables. In the case of the models based on raw

and normalized spectra data, the best correlations were provided by sixteen and nineteen CSF-1R inhibitor variable models, respectively, with variables being selected in association to wavenumbers that presented high PC1 and PC2 loading values. The wavenumbers selected for the final models were: 3163, 2970, 2916, 2847, 2212, 2033, 1906, 1802, 1553, 1152, 947, 918, 872, RG7204 clinical trial 841, 789 and 750 cm−1 (raw data); 3125, 2991, 2498, 2125, 1958, 1780, 1641, 1539, 1331, 1171, 1134, 978, 908, 864, 833, 808, 806, 754 and 725 cm−1 (normalized data). There were also several attempts of obtaining a model based on spectra derivatives, since this type of spectra manipulation was the most effective in providing separation between pure corn, coffee and coffee husks (see Fig. 4c). However, it was not possible to obtain a model that could provide satisfactory discrimination and thus only the models based on raw and normalized data will be presented. The developed model equations can be represented by: equation(1)

DFi=C0+∑j=1NCjAjwhere DFi represents the discriminant Farnesyltransferase function (i = 1,2,3), N is the total number of variables in the model, and Aj is the model variable, i.e., absorbance value at the selected wavenumber (model based on raw spectra data) or absorbance value at the selected wavenumber after normalization and baseline correction (Model based on normalized data). The corresponding model coefficients (Cj) are displayed in Table 2 and the score plots obtained for the three discriminant functions are shown in Fig. 5. The first two discriminant functions accounted for 84 and 91% of the total sample variance, for the models based on raw and normalized spectra, respectively. A clear separation between pure roasted coffee and roasted adulterants (coffee husks and corn), as well as adulterated coffee samples, can be observed for both models (see Fig. 5a and b). Notice that, for the adulterated samples, there is a wider dispersion of the data due to the differences in both the nature of the adulterants and their content in the adulterated samples. The calculated values of each discriminant function at the group centroids are displayed in Table 3.