Pectin, one of the most structurally complex families of polysaccharides in nature, is widely used as functional food ingredient. Pectin properties are affected by factors related to its structure so there is an overall interest in finding alternative sources of pectin with enhanced bioactivity. Artichoke pectin extraction has been described, however low correlation coefficients were obtained when modelling and predicting the extraction process, probably due to its complexity. Currently, artificial neural networks (ANN) are gaining great relevance, since it allows modelling complex and highly non-linear processes and have been recently used to determine structural patterns of artichoke pectic oligosaccharide GC-MS-EI spectra. Therefore, the aim of this work was to evaluate ANN as an alternative tool for modelling and optimising enzymatic artichoke pectin extraction process using a commercial cellulase, Celluclast®. The independent variables used in this study were artichoke by-product powder concentration (2-7% w/w), cellulase dose (2.2-13.3 U g-1), and extraction time (6-24 h), being the dependent variables pectin yield (mg 100 mg-1, range 9.0 - 19.1), weighted averaged Mw (kDa, range 189 - 344), galacturonic acid (GalA, mg 100 mg-1, range 1.07 - 2.20), pectic neutral sugars (mg 100 mg-1, range 0.71 - 2.29) and other sugars (mg 100 mg-1, range 0.47 - 1.51) contents. Then, different ANN were computed. These models are formed by an input layer (i.e. independent variables), an output layer (i.e. dependent variables) and several neurons or nodes organised in hidden layers. R2, R2 adj and root mean square error (RMSE) were used to evaluate the quality of the fit. Models were trained with 70% of data and tested on 30%, corresponding to new samples. According to the results obtained, High R2 and R2 adj and low RMSE values were obtained during the training phase for pectin yield (0.99, 0.98, 2.7%), weighted averaged Mw (0.97, 0.96, 4.9%), GalA (0.92, 0.91, 7.0%), pectic neutral sugars (0.98, 0.97, 3.6%) and other monosaccharides (0.97, 0.96, 3.8%). Similar values were observed during the test phase: pectin yield (0.97, 0.95, 6.3%), GalA (0.93, 0.90, 8.7%), pectic neutral sugars (0.99, 0.97, 3.1%) and contaminant monosaccharides (0.98, 0.96, 0.04%). However, a poor fit for Mw was obtained during this step (0.77, 0.65, 9.4%). These results highlight ANN high predictive power and its applicability for modelling functional ingredient extraction from industrial by-products of more effective form.

Resumen del trabajo presentado a las III Jornadas Científicas CIAL Fórum, celebradas del 22 al 23 de noviembre de 2018 en el Instituto de Investigación en Ciencias de la Alimentación (CIAL).

Projects AGL2014-53445-R and AGL2017-84614-C2-1-R (MICINN, Spain). Carlos Sabater thanks FPU Predoc contract (FPU14/03619, MECD, Spain).

Peer reviewed