SNP genotyping in Pseudotsuga menziesii and Pinus radiata using targeted genotyping-by-sequencing (GBS): improved Bayesian SNP calling using a beta-binomial distribution and other optimized input parameters.

backgroundSingle-nucleotide polymorphism markers (SNPs) have important applications in gene conservation, breeding, and fundamental genetics research. Our long-term goal is to develop routine approaches for SNP genotyping in forest trees. Ideally, these approaches would be inexpensive, able to accommodate a wide range of samples and SNPs, available through commercial providers, and produce high-quality SNP data.

resultsUsing targeted genotyping-by-sequencing (GBS), we developed SNP assays for two highly heterozygous tree species, Douglas-fir (Pseudotsuga menziesii) and radiata pine (Pinus radiata). Using Douglas-fir haploid and diploid data, we optimized Bayesian SNP calling by testing four input parameters: (1) allele and genotype prior probabilities, (2) Rho, the beta-binomial dispersion parameter, (3) estimated read error (BayesReadError), and (4) the logPO cutoff used to filter low confidence SNP calls. logPO is the Bayesian posterior odds ratio for a called SNP. Compared to assuming a binomial distribution of read counts (Rho = 0), the beta-binomial distribution (Rho = 0.33) substantially reduced call error and heterozygote undercalling. Compared to the other Bayesian parameters, genotype priors had little effect on genotyping success. For Douglas-fir, we tested 5,360 SNP assays, and then studied the performance of the best 4,000. For radiata pine, we tested 6,000 SNP assays, and then studied the performance of the best 4,570. In Douglas-fir and radiata pine, our Bayesian approach resulted in median call rates of 95% to 98% for the top-ranked SNPs, with an estimated call error of 1.60% for known homozygous genotypes and 2.27% for known heterozygotes. In radiata pine, median and mean call rates were above 91% for GBS and SNP genotyping using an Axiom fixed genotyping array. Additionally, the median correspondence between the GBS and Axiom genotypes was about 98% overall (mean 96%).

conclusionsBy optimizing Bayesian SNP calling, selecting the best 4-5 K SNPs, and excluding samples with low DNA amounts, we substantially reduced call error and heterozygote undercalling, resulting in SNP genotypes that were nearly identical to genotypes obtained using the Axiom array. Furthermore, genotyping performance should increase even further if our SNP rankings were used to develop less complex probe pools that target fewer SNPs.