Application of near infrared spectroscopy for green coffee biochemical phenotyping

Accessions resulting from surveys in Ethiopia (the centre of origin of Arabica coffee) can be used as a source of genetic variability in breeding coffee plants. They may contain some genes of interest for coffee breeding, specifically in relation to beverage quality. Near infrared (NIR) spectroscopy was used to develop models for predicting the major coffee constituents related to quality beverage (proteins, caffeine, lipids, chlorogenic acids, phenolic compounds, total sugars and sucrose). We selected coffee samples listed in a database containing data of chemical contents from samples of traditional and modern cultivars and of Ethiopian accessions to construct models to predict these compounds. Spectra were collected between 1100nm and 2500nm, and mathematical pretreatments were applied. The number of samples for the calibration step for each compound was set so as to be representative of distribution values. Cross-validation was performed on the total set of samples to select the optimal number of terms for the prediction models of each component. The prediction models were developed employing a modified partial least-squares regression. The total set of samples for each component was divided randomly into two subsets: one for developing the prediction model and the other to evaluate the predicted values. The best prediction models obtained were for chlorogenic acids (r2= 0.94, RPD=4.16), proteins (r2=0.94, RPD= 4.09) and caffeine (r2=0.92, RPD= 4.16) . Models for lipids and phenolic compounds were not as accurate (r2=0.87, RPD=2.77 and r2=0.86, RPD= 2.62, respectively), while models for sucrose (r2= 0.84, RPD= 2.44) and total sugars (r2=0.85, RPD= 2.55) were even less accurate. All these models can be used for identifying coffee lines with more desirable traits in breeding programmes. The models were effective in discriminating Ethiopian coffee accessions from modern cultivars of coffee. Additionally, the NIR technique will make it possible to analyse a large nu