A rapid method for the differentiation of yeast cells grown under carbon and nitrogen-limited conditions by means of partial least squares discriminant analysis employing infrared micro-spectroscopic data of entire yeast cells

Abstract

This paper shows the ease of application and usefulness of mid-IR measurements for the investigation of orthogonal cell states on the example of the analysis of Pichia pastoris cells. A rapid method for the discrimination of entire yeast cells grown under carbon and nitrogen-limited conditions based on the direct acquisition of mid-IR spectra and partial least squares discriminant analysis (PLS-DA) is described. The obtained PLS-DA model was extensively validated employing two different validation strategies: (i) statistical validation employing a method based on permutation testing and (ii) external validation splitting the available data into two independent sub-sets. The Variable Importance in Projection scores of the PLS-DA model provided deeper insight into the differences between the two investigated states. Hence, we demonstrate the feasibility of a method which uses IR spectra from intact cells that may be employed in a second step as an in-line tool in process development and process control along Quality by Design principles.

Graphical abstract

Fig. 1

Mean spectra of different data sets (raw data) and close-up view of the region between 1000 and 1100 cm⁻¹ (a) and mean spectra of different data sets after derivative (2nd derivative, points: 9, polynomial order: 3) and normalization (b). Note: CAL stands for calibration set and VAL stands for validation set (for details about the normalization see Experimental section).

Fig. 2

Scores plots of PC1 vs. PC2 (a), scores plots of PC2 vs. PC3 (b) Hotelling T² vs. Q residuals (c) obtained from PCA.

Fig. 3

Results of the PLS-DA permutation test using the calibration dataset (9 samples with 3 replicates each): number of misclassifications (a), Q² values (b), AUROC values (c) and sample predictions using the real class labels (d).

Fig. 4

Score plot of the class prediction of samples included in calibration (first 27 data points) and validation set (last 12 data points (a), Q Residuals vs. Hotelling T² values for calibration (CAL) and validation data (VAL) (b) and PLS-DA VIP Scores (c); Note: mean spectrum (×50) of the calibration data set has been included in (c) for an easier interpretation; for a better visibility it has been shifted in the y direction.