INVESTIGADORES
MAMMARELLA Enrique Jose
congresos y reuniones científicas
Título:
An artificial neural network approach to predict lactose hydrolysis by beta-galactosidase
Autor/es:
CUELLAS, ANAHÍ; MAMMARELLA, ENRIQUE; RUBIOLO, AMELIA; IAMMARINO, MARCELO; ODDONE, SEBASTIÁN; CORRADINI, MARÍA
Lugar:
Anaheim, CA, Estados Unidos
Reunión:
Otro; IFT Annual Meeting; 2009
Institución organizadora:
Institute of Food Technologists
Resumen:
The disposal of whey permeate remains a significant
environmental problem for the dairy industry because of its high lactose
content. Lactose transformation to value added products such as sweetening
syrups or ethanol constitutes a feasible strategy if conversion rates and yields
are optimized. The hydrolysis of lactose is a key step in any conversion process
and it can be accomplished using β-galactoside. The objective of this work was
to predict the performance of enzyme reactors at various operational conditions
using an artificial neural network (ANN) approach. The ANN technique enables us
to investigate the effects of the operational factors simultaneously instead of
evaluating their effects separately.
For that purpose, 280 experimental data sets were obtained using
several substrate concentrations (0.07, 0.15, 0.22, 0.29 M), enzyme
concentrations (0.08-3 mg/ml), reaction media (lactose solution, whey permeate
& whey permeate with proteins) and enzyme conditions (free vs. immobilized
covalently bound to chitosan beads). All experiments were carried out in
triplicate.
A back propagation three-layer neural network model was
developed. The input layer had six elements: time, enzyme condition, enzyme
concentration, initial substrate concentration, reaction media and protein
concentration. The number of neurons in the hidden layer was adjusted to a final
value of 13, and the results were analyzed for over- and under-fitting. The
output layer had only one element; the residual lactose concentration. The
experimental dataset was divided into two groups by random selection, the first
group (150 observations) was used for network training and the remaining data
(130 observations) was used for validation.