114 Accuracy of genomic predictions using single and multiple-trait machine learning methods in Canadian beef cattle population

Abstract The advancement in various machine learning (ML) methods provides tools to extract features from a large data set of complex traits and to predict the outcome of the target trait. In this study, we assessed the genomic prediction accuracy using single trait models based on kernel ridge regression (ST_KRR) and linear support vector regression (STLinearSVR) for female feed intake, feed efficiency and fertility traits of Canadian crossbreed beef cattle (n = 2,834 genotyped with 83,875 single nucleotide polymorphisms) and compared their prediction performances to a single trait genomic best linear unbiased prediction (STGBLUP) method. We also evaluated the performances of multiple-trait genomic prediction methods including MAK-based KRR, MAK-based LinearSVR (MAK_MTLinearSVR), and multiple-trait genomic best linear unbiased prediction (MTGBLUP). For the single trait methods, we found that the ST_KRR model was 1.86% to 26.86% more accurate for feed intake and fertility traits including residual feed intake adjusted for backfat thickness (RFIfat), daily dry matter intake (DMI), feeding event duration (DUR), age of first calving (AFC), pre-breeding backfat at first parity (PBBF), and pre-breeding weight at first parity (PBWT) than the STGBLUP method, which had an average accuracy of 0.39 for the traits investigated. For growth and development traits including julian date of birth (JulianDt), weaning weight adjusted to 200 d (WT200d), and birth weight (BRWT) of heifers, STLinearSVR yielded a 5.44% to 9.92% greater accuracy than STGBLUP. As for the multiple trait methods, MAK_MTLinearSVR had a 1.1% to 20.42% greater accuracy for PBWT, AFC, WT200d, DUR, RFIfat, and BRWT than the single trait model of STLinearSVR, while MTGBLUP had a 0.42% to 32.64% greater accuracy for AFC, RFIfat, WT200d and PBBF than the single trait model of STGBLUP. Our results demonstrate the potential application of ML-based multiple-trait models in beef cattle to improve the effectiveness of genomic selection programs, in particular for traits with low heritability.