Assessment of the Structure and Predictive Ability of Models Developed for Monitoring Key Analytes in a Submerged Fungal Bioprocess Using Near-Infrared Spectroscopy

Abstract

The robustness of models developed for the near-infrared spectroscopic prediction of mycelial biomass, total sugars, and ammonium in a submerged Penicillium chrysogenum bioprocess was assessed by rigorously challenging them with artificially introduced analyte and background matrix variations, so that analyte concentrations were varied in an invariant matrix and vice versa. The models were also challenged by using a data set from a process operated at a different scale from that used in the original model formulation. Simple univariate and bivariate linear regression models, and partial least-squares (PLS) models with as few factors as three and four, performed sufficiently well for predicting analyte concentrations and were robust with respect to the matrix variations tested. However, models based on relatively weaker absorptions, or those that were likely to be influenced by stronger absorbers present in the same matrix, were vulnerable to changes in the matrix. A change in the scale of operation affected models that would be influenced by biomass, possibly due to an influence of the morphology of the mycelial biomass. An analysis of the loading vectors of some PLS models revealed details that were useful in understanding the type of information modeled and the behavior of these models to the variations tested.