Quantification of individual fatty acids in bovine milk by infrared spectroscopy and chemometrics: Understanding predictions of highly collinear reference variables

Predicting individual fatty acids (FA) in bovine milk from Fourier transform infrared (FT-IR) measurements is desirable. However, such predictions may rely on covariance structures among individual FA and total fat content. These covariance structures may change with factors such as breed and feed, among others. The aim of this study was to estimate how spectral variation associated with total fat content and breed contributes to predictions of individual FA. This study comprised 890 bovine milk samples from 2 breeds (455 Holstein and 435 Jersey). Holstein samples were collected from 20 Danish dairy herds from October to December 2009; Jersey samples were collected from 22 Danish dairy herds from February to April 2010. All samples were from conventional herds and taken while cows were housed. Moreover, in a spiking experiment, FA (C14:0, C16:0, and C18:1 cis-9) were added (spiked) to a background of commercial skim milk to determine whether signals specific to those individual FA could be obtained from the FT-IR measurements. This study demonstrated that variation associated with total fat content and breed was responsible for successful FT-IR–based predictions of FA in the raw milk samples. This was confirmed in the spiking experiment, which showed that signals specific to individual FA could not be identified in FT-IR measurements when several FA were present in the same mixture. Hence, predicted concentrations of individual FA in milk rely on covariance structures with total fat content rather than absorption bands directly associated with individual FA. If covariance structures between FA and total fat used to calibrate partial least squares (PLS) models are not conserved in future samples, these samples will show incorrect and biased FA predictions. This was demonstrated by using samples of one breed to calibrate and samples of the other breed to validate PLS models for individual FA. The 2 breeds had different covariance structures between individual FA and total fat content. The results showed that the validation samples yielded biased predictions. This may limit the usefulness of FT-IR–based predictions of individual FA in milk recording as indirect covariance structures in the calibration set must be valid for future samples. Otherwise, future samples will show incorrect predictions.