Predicting milk protein fractions using infrared spectroscopy and a gradient boosting machine for breeding purposes in Holstein cattle

• DOI: 10.3168/jds.2022-22119
• Publication Date: 2023-01-28T02:21:07Z
• Authors: L.F. Macedo Mota; V. Bisutti; A. Vanzin; S. Pegolo; A. Toscano; S. Schiavon; F. Tagliapietra; L. Gallo; P. Ajmone Marsan; A. Cecchinato
• Keywords: Boosting; Holstein Cattle

In recent years, increasing attention has been focused on the genetic evaluation of protein fractions in cow milk with aim improving quality and technological characteristics. this context, advances high-throughput phenotyping by Fourier transform infrared (FTIR) spectroscopy offer opportunity for large-scale, efficient measurement novel traits that can be exploited breeding programs as indicator traits. We took samples from 2,558 Holstein cows belonging to 38 herds northern Italy, operating under different production systems. spectra were collected same day sampling stored subsequent analysis. Two sets data (i.e., phenotypes FTIR spectra) 2 years (2013 2019–2020) compiled. The following assessed using HPLC: true protein, major casein [αS1-casein (CN), αS2-CN, β-CN, κ-CN, glycosylated-κ-CN], whey proteins (β-lactoglobulin α-lactalbumin), all which measured both grams per liter (g/L) proportion total nitrogen (% N). predictions calculated gradient boosting machine technique tested 3 cross-validation (CRV) methods. used CRV scenarios: (1) random 10-fold, randomly split whole into 10-folds equal size (9-folds training 1-fold validation); (2) herd/date-out CRV, assigned 80% herd/date set independence 20% validation set; (3) forward/backward according year (FTIR gold standard 2013 or "old" "new" databases validation, vice-versa among them; (4) parameters (CRV-gen), where animals without pedigree a fixed population information was 5-folds, population, 4-folds (independent set). results measures predictions) CRV-gen infer laboratory measurements HPLC) FTIR-based considering scenario bi-trait animal model single-step genomic BLUP. found prediction accuracies equations differed way expressed, achieving higher accuracy when expressed g/L than % N scenarios. Concerning reproducibility over showed no relevant differences predictive ability between vice-versa. Comparing additive variance estimates predicted HPLC measures, we reductions −19.7% g/L, −21.19% N. Although heritability estimates, they small, values ranging −1.9 −7.25% −1.6 −7.9% posterior distributions correlations (ra) generally high (>0.8), even Our show potential selection enhance fraction contents. expect acceptable responses due predictions.