A Fast and Flexible Method for Improving Genomic Prediction with Biological Information

Genomic prediction has emerged as an effective approach in plant and animal breeding and in precision medicine. Much research has been devoted to achieve higher accuracy in genomic prediction, and one of the potential ways is to incorporate biological information. Due to the statistical and computational challenges, however, a fast and flexible method to incorporate such external information is still lacking. Here, we propose a linear mixed model that can incorporate biological information in a flexible way and develop a fast variational Bayes-based software package. In our proposed model, whole genome markers (m in total) are split into p (1≤p≤m) groups in a user-defined manner, and the kth group of markers is given a common effect variance (k = 1,…,p). Additionally, each marker has a pre-specified weight for which the rule can be flexibly defined. We provide examples on using the model to incorporate biological information. Previous functional genomics studies have accumulated much evidence on which genes or pathways are more/less important for milk yield in dairy cattle. Based on their levels of importance, we can divide genome-wide markers into a number of groups in our model. We also know rare variants are more likely to be QTLs than common variants for some traits, so we may define a rule, ω_j=Beta(MAF_j;α=1,β=10) (beta distribution density function), to pre-specify weight for the jth marker. Our proposed model is implemented with the parameter expanded variational Bayesian method, making it fast and feasible to analyze very large data sets. The software is written in C++ with the Intel MKL library. As a test, we analyzed a large cattle data set consisting of ~20k old bulls (training population) and ~4k young bulls (validation population). All animals have genotyped or imputed ~760k whole-genome sequence genotypes. By grouping markers based on proximity, our software performed better than the industrial implementation of Bayes A in all five milk traits analyzed, with an increase of up to 10% in prediction accuracy and a similar running time. Collectively, the method and software show great potential to increase accuracy in genomic prediction, particularly in the future when more useful biological information is becoming available.