[en] Mid-Infrared spectrometry predicts the milk components (e.g., %fat, %protein) from spectral data reecting the milk composition. The data included 9,663 test days on 1,937 cows in 1 to 12 parity recorded from April 2005 to May 2006. Each sample was scanned by MilkoScan FT6000 into 1,060 points. Due to the high dimension, principal components approach (PCA) was done to reduce the traits and indicated that 48 principal components (PC) described 99.02% of information. These PC were analyzed by multi-trait REML using the canonical transformation. This analysis considered 2,850 rst lactation records for 738 cows in 7 breeds from 26 herds. Effects included in the multi-trait model were herd*test date, lactation stage, permanent environmental and animal random effects. The estimates of the variances were back transformed to the initial scales. Heritabilities varied from 0.005% to 57.20% for the different pin numbers. Spectral regions with heritability greater than 5% were located between 1 to 181; 194 to 558 and 709 to 1,060 pin numbers. PCA involving points in those regions demonstrated that only 9 PC explained 99.23% of information. MidInfrared spectrum contains specic regions with substantial genetic information potentially useful for selecting improved milk quality directly on spectral data.