Comparison of methods for wavelength combination selection from multi-wavelength fluorescence spectra for on-line monitoring of yeast cultivations.

Abstract

The on-line monitoring with two-dimensional (2D) fluorescence spectroscopy of Saccharomyces cerevisiae batch cultivations was applied to monitor glucose, ethanol, and biomass concentrations. The measurement of one spectrum by the 2D fluorescence spectrometer has 120 fluorescence intensity values of excitation and emission wavelength combinations (WLCs); scattered light is not considered here. To identify which WLCs of the multi-wavelength fluorescence spectra carry important and relevant information regarding the analyte concentrations, three different methods were compared: a method based on loadings, variable importance in projection, and ant colony optimization. The five selected WLCs for a particular analyte from each method were evaluated by multiple linear regression models to find the most significant variable subsets for predicting the sample concentrations. The most significant WLCs relevant to the three sample properties contained seven different excitation and emission wavelengths, which can combine with each other to have 38 possible wavelength combinations in the fluorescence measurement. Partial least squares (PLS) models were calibrated with the 38 possible variables and the off-line data for the prediction of glucose, ethanol, and biomass concentrations. The best prediction from the PLS models had the percentage of root mean square error of prediction (pRMSEP) in the range of 3.1-6.3%, which was similar to pRMSEPs of the PLS models with the full variables. Based on these results, it is promising to build up a specific inexpensive fluorescence sensor for the yeast cultivation process using light-emitting diodes and photodiodes. Graphical abstract Flow chart of the selection of important and relevant WLCs by the three methods to predict glucose, ethanol, and biomass concentrations.