Ranking Individuals Based On Predicted Performance Using Mean Normalized Discounted Cumulative Gain Value

Abstract

Accuracy of genomic prediction always relies on an appropriate choice of a statistical model to capture the relationship between the genetic architecture of a trait and the underlying marker calls in a panel of high-density marker data. However, the ranking problem has become an important subject in machine learning (ML) models, due to its widespread applications in many decision-making processes because the measures of rank quality are usually based on sorting, which is not directly optimizable. To counter this, mean normalized discounted cumulative gain value (MNDCGV), a standard quality measure in information retrieval with capabilities of ranking individuals according to breeding values has been proposed. Few studies have emphasized on the ranking of individuals based on predicted phenotypic values using MNDCGVs but none have been reported in animals. The focus of this study, therefore, was, to evaluate the prediction performance of DeepGS, RR-BLUP and Ensemble GS models using MNDCGV. The MNDCGV results showed the accuracy of GEBVs estimated using DeepGS was approximately equal to 0.75~0.78, RR-BLUP 0.66~0.76 and Ensemble 0.76~0.79 as a result of top-ranked alpha increasing from 10% to 70%. The Ensemble and DeepGS model outperformed the conventional RR-BLUP model by a significant margin (P<0.05), therefore they can be used as a supplement to RR- BLUP. Thus, Ensemble and DeepGS models can be given a top priority as GS model and as an alternative to conventional GS models in predicting the performance and ranking of individuals with high breeding values to be used for selection purposes in indigenous chicken breeding programs. Ensemble model performed very well in ranking individuals with better performance compared to DeepGS and RR-BLUP, with improvement values of 0.01 and 0.11, respectively. Thus, Ensemble model can be given a top priority as GS model for performance improvement.