Quantile regression for prediction of complex traits in Braunvieh cattle using SNP markers and pedigree

Abstract

Genomic prediction models generally assume that errors are distributed as normal, independent, and identically distributed random variables with zero mean and equal variance. This is not always true, in addition there may be phenotypes distant from the population mean, which are usually removed when making the prediction. Quantile regression (QR) deals with statistical aspects such as high dimensionality, multicollinearity and phenotypic distribution different from the normal one. The objective of this work was to compare QR using marker and pedigree information with alternative methods such as genomic best linear unbiased prediction (GBLUP) and single-step genomic best linear unbiased prediction (ssGBLUP) to analyze the birth (BW), weaning (WW) and yearling (YW) weights of Braunvieh cattle and simulated data with different degrees of asymmetry and proportion of outliers. The predictive capacity of the models was assessed by cross-validation. The predictive performance of QR both with marker information alone and with information of markers plus pedigree, with the actual dataset, was better than the GBLUP and ssGBLUP methodologies for WW and YW. For BW, GBLUP and ssGBLUP were better, however, only quantiles 0.25, 0.50 and 0.75 were used, and the BW distribution was not asymmetric. In the simulated data experiment, correlations between “true” marker effects and estimated effects, as well as “true” and estimated signal correlations were higher when QR was used compared to GBLUP. The advantages of QR were more noticeable with asymmetric distribution of phenotypes and with a higher proportion of outliers, as was the case with the simulated dataset.