Prediction of milk fatty acid content with mid-infrared spectroscopy in Canadian dairy cattle using differently distributed model development sets

The fatty acid profile of milk is a prevailing issue due to the potential negative or positive effects of different fatty acids to human health and nutrition. Mid-infrared spectroscopy can be used to obtain predictions of otherwise costly fatty acid phenotypes in a widespread and rapid manner. The objective of this study was to evaluate the prediction of fatty acid content for the Canadian dairy cattle population from mid-infrared spectral data and to compare the results produced by altering the partial least squares (PLS) model development set used. The PLS model development sets used to develop the predictions were reference fatty acids expressed as (1) grams per 100 g of fatty acid, (2) grams per 100 g of milk, (3) the natural logarithmic transform of grams per 100 g of milk, and (4) subsets of samples randomly selected by removing excess records around the mean to present a more uniform distribution, repeated 10 times. Gas chromatography measured fatty acid concentration and spectral data for 2,023 milk samples of 373 cows from 4 breeds and 44 herds were used in the model development. The coefficient of determination of cross-validation (Rcv2) increased when fatty acids were expressed on a per 100 g of milk basis compared with on a per 100 g of fat basis for all examined fatty acids. The logarithmic transformation used to create a more Gaussian distribution in the development set had little effect on the prediction accuracy. The individual fatty acids C12:0, C14:0, C16:0, C18:0, C18:1n-9 cis, and saturated, monounsaturated, unsaturated, short-chain, medium-chain, and long-chain fatty acid groups had (Rcv2) greater than 0.70. When model development was performed with subsets of the original samples, slight increases in (Rcv2) values were observed for the majority of fatty acids. The difference in (Rcv2) between the top- and bottom-performing prediction equation across the different subsets for a single predicted fatty acid was on average 0.055 depending on which samples were randomly selected to be used in the PLS model development set. Predictions for fatty acids with high accuracies can be used to monitor fatty acid contents for cows in milk recording programs and possibly for genetic evaluation.