Prediction of fatty acid chain length and unsaturation of milk fat by mid-infrared milk analysis1

Our objective was to develop partial least squares (PLS) models to predict fatty acid chain length and total unsaturation of milk fat directly from a mid-infrared (MIR) spectra of milk at 40°C and then determine the feasibility of using those measures as correction factors to improve the accuracy of milk fat determination. A set of 268 milks (modified milks, farm bulk tank milks, and individual cow) were analyzed for fat, true protein, and anhydrous lactose with chemical reference methods, and in addition a MIR absorption spectra was collected for each milk. Fat was extracted from another portion of each milk, the fat was saponified to produce free fatty acids, and the free fatty acids were converted to methyl esters and quantified using gas-liquid chromatography. The PLS models for predicting the average chain length (carbons per fatty acid) and unsaturation (double bonds per fatty acid) of fatty acids in the fat portion of a milk sample from a MIR milk spectra were developed and validated. The validation performance of the prediction model for chain length and unsaturation had a relative standard deviation of 0.43 and 3.3%, respectively. These measures are unique in that they are fat concentration independent characteristics of fat structure that were predicted directly with transmission MIR analysis of milk. Next, the real-time data output from the MIR spectrophotometer for fatty acid chain length and unsaturation of milk were used to correct the fat A (C=O stretch) and fat B (C-H stretch) measures to improve accuracy of fat prediction. The accuracy validation was done over a period of 5 mo with 12 sets of 10 individual farm milks that were not a part of the PLS modeling population. The correction of a traditional fat B virtual filter result (C-H stretch) for sample-to-sample variation in unsaturation reduced the Euclidean distance for predicted fat from 0.034 to 0.025. The correction of a traditional fat A virtual filter result (C=O stretch) modified with additional information on sample-to-sample variation of chain length and unsaturation gave the largest improvement (reduced Euclidean distance from 0.072 to 0.016) and the best validation accuracy (i.e., lowest Euclidean distance) of all the fat prediction methods.