Genotype imputation and error estimation in connected multiparental populations

Abstract Multiparental populations have been produced for quantitative trait loci mapping in many crops, where next-generation sequencing has become a cost-effective tool for genotyping. Previously, we have developed a hidden Markov framework denoted by MagicImpute_mma for genotype imputation in a multiparental population, which was implemented in Mathematica. However, its computational time increases quickly with the number of founders. In this work, we extend MagicImpute_mma into MagicImpute for increasing computational efficiency and robustness to various types of errors. Particularly, it has the following novel features: (i) allowing for multiple multiparental populations that may be connected by sharing founders, (ii) allowing for many missing founders that are not available for sequencing, (iii) accounting for allelic bias and overdispersion in next-generation sequencing data, (iv) inferring marker-specific error rates and filtering for markers with low error rates, and (v) being implemented in the high-performance Julia language. Besides extensive simulation studies, we evaluate MagicImpute by 3 real datasets: the rice F2 population with sequence depth being low, the apple cross-pollinated population with parents being outbred, and the sorghum multiparent advanced generation intercross population with 10 male sterile lines (out of 29 founders) missing. The results show that MagicImpute is well capable of imputing many missing founders if there are sufficient offspring (eg 20 per founder) and that MagicImpute is accurate for genotype imputation in connected bi- or multiparental populations with various types of sequence errors, and it opens up new opportunities for quantitative trait loci mapping after imputing many missing founders.