Non Linear Modelling For Genomic Predictions Based on Multiple Traits

Abstract

Genomic Selection is a modern technique used to improve the genetic gain through breeding program. Genomic selection can be done both on single trait and multi trait data. Multi-trait genomic selection is better than single-trait genomic selection if genetic correlation between the traits is observed. However, multi-trait genomic selection has little been explored in the view of non-linearity problem between responses (traits) and explanatory variables (markers). In this study, problem of non-linear relation between markers and traits has been taken care in Multi Trait Genomic Selection (MTGS) framework. First of all, a comprehensive comparison between methods of single trait and multi-trait genomic selection has been made by using real data set. It has been observed that MTGS methods are better than STGS methods. In Genomic Selection, number of markers are very large which creates the large p and small n problem (n<<p). Many feature selection techniques are proposed to select the important markers associated with the traits. Among these techniques penalized regression like Least Absolute Shrinkage and Selection Operator (LASSO), Elastic Net etc. are most popular as they deal with n<p problem. But being a linear technique, they are unable to handle the problem of non-linear input-output dependency. Kernelised LASSO resolves the issues of n<<p as well as non-linearity problem simultaneously. We have used this technique in Multi-trait Genomic Selection (MTGS) and demonstrated it with real genomic and phenotypic data. Performance of the proposed method has been found to be fairly impressive. We have also developed web based software which has been named as “Genomic Selection Tool (GST)”. This on-line tool has been implemented the proposed method for prediction of genome estimated breeding value (GEBVs). This tool requires genotypic and phenotypic data belongs to different traits. This tool, predicts genome estimated breeding values for single trait as well as multi-traits. We have incorporated some of existing single trait and multi-traits methods along with our proposed method in this tool. GEBVs predicted by this tool may provide useful postulate for breeders for selection of important variety or breed.