MOUSE GENETIC RESOURCES Imputation of Single-Nucleotide Polymorphisms in Inbred Mice Using Local Phylogeny

We present full-genome genotype imputations for 100 classical laboratory mouse strains, using a novel method. Using genotypes at 549,683 SNP loci obtained with the Mouse Diversity Array, we partitioned the genome of 100 mouse strains into 40,647 intervals that exhibit no evidence of historical recombination. For each of these intervals we inferred a local phylogenetic tree. We combined these data with 12 million loci with sequence variations recently discovered by whole-genome sequencing in a common subset of 12 classical laboratory strains. For each phylogenetic tree we identified strains sharing a leaf node with one or more of the sequenced strains. We then imputed highand medium-confidence genotypes for each of 88 nonsequenced genomes. Among inbred strains, we imputed 92% of SNPs genome-wide, with 71% in high-confidence regions. Our method produced 977 million new genotypes with an estimated per-SNP error rate of 0.083% in high-confidence regions and 0.37% genome-wide. Our analysis identified which of the 88 nonsequenced strains would be the most informative for improving full-genome imputation, as well as which additional strain sequences will reveal more new genetic variants. Imputed sequences and quality scores can be downloaded and visualized online. AMONG the many advantages of inbred strains in genetic studies is that each strain needs to be genotyped only once, and that information can be reused in many experiments. Moreover, as more genotype data become available for a given inbred strain, the analysis can be updated. This cycle can continue until, ultimately, all inbred strains are fully sequenced. In the meantime, there is a need to leverage the handful of inbred strains that have been sequenced using robust imputation methods to maximize the value of existing data. High-quality imputed sequence has many potential applications including identification of functional variants and the creation of accurate scaffolds for the analysis of next-generation RNAseq and bisulfite sequencing data. Until affordable deep sequencing becomes a reality, a balanced approach that combines targeted sequencing with accurate imputation offers the best of both worlds: high-quality genomic data today at little additional cost. A recent sequencing effort by the Wellcome Trust/Sanger Institute has made available dense genome sequences for a set of 17 inbred mouse strains, including 13 common laboratory strains, 3 wild-derived mouse strains from different subspecies of Mus musculus, and a single strain from a different species, M. spretus (Keane et al. 2011). This set of samples is expected to capture much of the variation found in common laboratory mouse strains and, therefore, provides a foundation for sequence imputation. A complementary resource is the recent release of Mouse Diversity Array (MDA) genotypes from 162 mouse strains (Yang et al. 2011). MDA is a high-density DNA microarray designed to assay diversity among commonly used laboratory mice (Yang et al. 2009). The density of SNP genotypes available on the MDA exceeds the density of recombination events accumulated over the development of the classical inbred strains and as such the MDA SNPs can provide a framework for imputation of the underlying whole-genome sequence. Imputation can be used to increase the effective resolution of a lower-density SNP panel to match that of a higherCopyright © 2012 by the Genetics Society of America doi: 10.1534/genetics.111.132381 Manuscript received July 10, 2011; accepted for publication October 20, 2011 Supporting information is available online at http://www.genetics.org/lookup/suppl/ doi:10.1534/genetics.111.132381/-/DC1. Corresponding author: Department of Computer Science, 319 Sitterson Hall, University of North Carolina, Campus Box 3175, Chapel Hill, NC 27599. E-mail: [email protected] Genetics, Vol. 190, 449–458 February 2012 449 density panel when there is a subset of samples common to both sets. Previous imputation methods use variations of a hidden Markov model (HMM) to infer sequence similarities and likely transitions between haplotypes. These methods employ probabilistic models based on local sequence similarity to infer the state of missing genotypes. Missing genotypes arise from two sources. No-calls (N’s) can indicate either technical noise or an unexpected sequence variant such as a nearby SNP or an indel that interferes with probe hybridization. A second, and more extensive, source of missing genotypes is due to differences in the density of marker sets between platforms. There have been two recent imputation efforts in the laboratory mouse (Szatkiewicz et al. 2008; Kirby et al. 2010). Szatkiewicz and co-workers imputed genotypes at 7.9 million loci by combining low-density genotypes from 51 classical and wild-derived inbred mouse strains with high-density SNP discovery data obtained on a subset of 16 inbred strains (Frazer et al. 2007; Yang et al. 2007). The authors imputed each locus consecutively across the genome, using an HMM to predict the most likely genotype among the possible alleles. Using this locus-by-locus method, they reported a 10.4% error rate over the entire genome and 4.4% error in high-confidence regions. High-confidence regions are defined by high posterior probability and cover 71% of the genome. Kirby and co-workers imputed genotypes in 94 classical and wild-derived laboratory strains for 8.27 million SNP loci reported in the National Institute of Environmental Health Sciences (NIEHS)/Perlegen set (Frazer et al. 2007), using expectation-maximized integrative imputation (EMINIM) (Kang et al. 2010), a different HMM method that predicts genotypes by estimating haplotype blocks from the smaller set of samples with high-density genotypes. The hidden states in their model correspond to the 16 NIEHS/Perlegen strains or a 17th unknown state. Their method models recombination between haplotype blocks rather than transitions between SNPs. The authors imputed 657 million genotypes with a reported error rate of 2.4% over the entire genome and 0.27% in regions with high confidence based on posterior probability in the HMM. These two methods do not explicitly take advantage of the local phylogenetic relationships present in classical inbred strains. This shortcoming is particularly significant given the strong population structure and the limited amount of haplotype diversity present in classical laboratory strains (Yang et al. 2011). Our approach estimates both haplotype blocks and the relatedness between them in the form of a local phylogenetic tree. In contrast to previous methods, our haplotype blocks and trees are inferred from a larger set of genotypes. This has the advantage that the larger set of samples can capture haplotype diversity that is not sampled in the smaller high-density set. The success of this approach requires that the SNP density in the larger sample set is sufficient to detect haplotype blocks, which is the case for the MDA genotypes (Yang et al. 2011). Moreover, the trees provide a measure of difference between haplotypes that is consistent with their evolutionary history. Here we report the combined use of the MDA and Wellcome Trust/Sanger Institute resources to impute the genotypes of 88 common inbred strains (Supporting information, Figure S1). Our approach takes advantage of the local phylogenetic relationships among inbred strains to determine the confidence of local imputation and is highly accurate over most of the genome of common strains. On the basis of this imputation we discuss strategies for future sequencing and SNP discovery in the laboratory mice and the efficient use of this resource for association studies. Imputed genotypes and imputation confidence are provided in Table S1 and use of these data should cite this article as a reference. They are also publicly available at http:// www.csbio.unc.edu/imputation/. Materials and Methods