Improving imputation quality in BEAGLE for crop and livestock data

Abstract

Imputation is one of the key steps in the preprocessing and quality control protocol of any genetic study. Most imputation algorithms were originally developed for the use in human genetics and thus are optimized for a high level of genetic diversity. Different versions of BEAGLE were evaluated on genetic datasets of doubled haploids of two European maize landraces, a commercial breeding line and a diversity panel in chicken, respectively, with different levels of genetic diversity and structure which can be taken into account in BEAGLE by parameter tuning. Especially for phasing BEAGLE 5.0 outperformed the newest version (5.1) which in turn also lead to improved imputation. Earlier versions were far more dependent on the adaption of parameters in all our tests. For all versions, the parameter ne (effective population size) had a major effect on the error rate for imputation of ungenotyped markers, reducing error rates by up to 98.5%. Further improvement was obtained by tuning of the parameters affecting the structure of the haplotype cluster that is used to initialize the underlying Hidden Markov Model of BEAGLE. The number of markers with extremely high error rates for the maize datasets were more than halved by the use of a flint reference genome (F7, PE0075 etc.) instead of the commonly used B73. On average, error rates for imputation of ungenotyped markers were reduced by 8.5% by excluding genetically distant individuals from the reference panel for the chicken diversity panel. To optimize imputation accuracy one has to find a balance between representing as much of the genetic diversity as possible while avoiding the introduction of noise by including genetically distant individuals. KEYWORDSimputation BEAGLE reference panel reference genome 1 23 special tools for both cases have been developed. As fully homozy-24 gous lines are commonly present in crops, the software TASSEL 25 (Bradbury et al. 2007) was developed to work well on this data 26 structure (Swarts et al. 2014). Since pedigrees in animal breeding 27 1 INVESTIGATIONS can be much denser than in human populations (both w.r.t. depth 28 and family size), tools like FImpute (Sargolzaei et al. 2014) and 29 AlphaImpute (Hickey et al. 2011) have been developed to fully 30 utilize this information. 31In the imputation process all those methods use the fact that physi-32 cally close markers are likely inherited together, resulting in non-33 random associations of alleles. These methods thereby rely on the 34 knowledge of the physical position or at least the order of markers 35 for modeling linkage and thus the resulting linkage disequilibrium 36 (LD). In contrast, the software LinkImpute (Money et al. 2015) ac-37 counts for LD between pairs of markers and not their physical 38 positions. This can be particularly relevant for species in which no 39 reference sequence is available or whose genomes are known for a 40 high amount of translocations and inversions. 41 In contrast to other methods using a HMM, the Markov chain in 42 BEAGLE is not initialized by ...

Accounting for epistasis improves genomic prediction of phenotypes with univariate and bivariate models across environments

“…Genotype calls were coded according to the allele counts of the B73 AGPv4 reference sequence (Jiao et al, 2017) (0 = homozygous for the reference allele, 2 = homozygous for the alternative allele). Imputation of missing values was performed separately for each landrace, using BEAGLE version 4.0 with parameters buildwindow=50, nsamples=50 (Browning and Browning, 2007;Pook et al, 2020). As the dataset only included doubled haploid lines and heterozygous calls were not expected, the DS (dosage) information of the BEAGLE output was used to recode remaining heterozygous calls.…”

HaploBlocker: Creation of Subgroup-Specific Haplotype Blocks and Libraries

“…Overall, the opportunities for identifying long, shared segments will be higher in SNP datasets from populations subjected to a recent history of intensive selection, as is commonly present in livestock and crop datasets. Recent work has suggested that the phasing accuracy for these kinds of datasets is extremely high (Pook et al 2019) and should, therefore, be sufficient for the application of HaploBlocker. For datasets containing less related individuals, as commonly present in human data, poor phasing accuracy can limit the applicability and usefulness of HaploBlocker.…”

Genome-wide scans for signatures of selection in Mangalarga Marchador horses using high-throughput SNP genotyping

“…However, in the ROH studies we chose to adopt an extreme lower parameter due to the presence of studies which recommended not using the MAF threshold (possible underestimation) [51]. For the iHS and TD analyses, the database was computed on Beagle version 5.0, which provides faster and more accurate algorithms for genotype haplotyping/phasing [52].…”