Fast genomic prediction of breeding values using parallel Markov chain Monte Carlo with convergence diagnosis

Peng Guo,Yonghu Liang,Lingyang Xu,Huijiang Gao,Xue Gao,Yong Guo,Xiaolin Wu,Junya Li,Lupei Zhang,Zezhao Wang,El Hamidi A Hay,Hong Niu,Yan Chen,Hemin Ni,Bo Zhu

doi:10.1186/s12859-017-2003-3

Abstract

BackgroundRunning multiple-chain Markov Chain Monte Carlo (MCMC) provides an efficient parallel computing method for complex Bayesian models, although the efficiency of the approach critically depends on the length of the non-parallelizable burn-in period, for which all simulated data are discarded. In practice, this burn-in period is set arbitrarily and often leads to the performance of far more iterations than required. In addition, the accuracy of genomic predictions does not improve after the MCMC reaches equilibrium.ResultsAutomatic tuning of the burn-in length for running multiple-chain MCMC was proposed in the context of genomic predictions using BayesA and BayesCπ models. The performance of parallel computing versus sequential computing and tunable burn-in MCMC versus fixed burn-in MCMC was assessed using simulation data sets as well by applying these methods to genomic predictions of a Chinese Simmental beef cattle population. The results showed that tunable burn-in parallel MCMC had greater speedups than fixed burn-in parallel MCMC, and both had greater speedups relative to sequential (single-chain) MCMC. Nevertheless, genomic estimated breeding values (GEBVs) and genomic prediction accuracies were highly comparable between the various computing approaches. When applied to the genomic predictions of four quantitative traits in a Chinese Simmental population of 1217 beef cattle genotyped by an Illumina Bovine 770 K SNP BeadChip, tunable burn-in multiple-chain BayesCπ (TBM-BayesCπ) outperformed tunable burn-in multiple-chain BayesCπ (TBM-BayesA) and Genomic Best Linear Unbiased Prediction (GBLUP) in terms of the prediction accuracy, although the differences were not necessarily caused by computational factors and could have been intrinsic to the statistical models per se.ConclusionsAutomatically tunable burn-in multiple-chain MCMC provides an accurate and cost-effective tool for high-performance computing of Bayesian genomic prediction models, and this algorithm is generally applicable to high-performance computing of any complex Bayesian statistical model.

Highlights

Running multiple-chain Markov Chain Monte Carlo (MCMC) provides an efficient parallel computing method for complex Bayesian models, the efficiency of the approach critically depends on the length of the non-parallelizable burn-in period, for which all simulated data are discarded
We noted that the loss of parallel computing efficiency relative to an assumedly perfect speedup increased with the model dimension, which is proportional to the number of SNPs in the genomic prediction models (Fig. 1)
Parallel computing of Bayesian genomic prediction models: tunable burn-in versus fixed burn-in Bayesian regression models are of high value for genomic prediction, the complexity of computing of these models can be intensive [14], which is increasingly becoming the bottleneck in practical genomic selection programs

Summary

Introduction

Running multiple-chain Markov Chain Monte Carlo (MCMC) provides an efficient parallel computing method for complex Bayesian models, the efficiency of the approach critically depends on the length of the non-parallelizable burn-in period, for which all simulated data are discarded. This burn-in period is set arbitrarily and often leads to the performance of far more iterations than required. Genomic predictions have been proposed as a method of providing accurate estimates of the genetic merits of breeding animals using genome-wide SNP markers [1] This new technology does not require the actual phenotyping of breeding candidates and offers great promise for traits that are difficult or expensive to measure, such as carcass traits [2]. Wu et al proposed the use of a multiple-chain MCMC method to calculate Bayesian genomic prediction models [14]

Methods

Results

Discussion

Conclusion