Pedigree-Based Genotype Imputation at Scale

Genotypic data drives many critical decisions that advance products through Monsanto’s breeding pipelines. Examples of genetically driven processes include genome-wide selection (GWS), marker-assisted selection (MAS), and identity by descent (IBD). High quality genotype estimates for missing data can greatly improve the accuracy of decisions and improve resource utilization. Common approaches for genotype imputation in the literature rely primarily on population-derived reference maps, assuming that pedigree information is shallow or incomplete. In agriculture, however, it is not uncommon to encounter genetic data, at various densities, connected through a large, well-annotated ancestry tree.

We describe a general-purpose imputation strategy that prioritizes utilization of pedigree information along with genotypic observations. The algorithm estimates probabilistic genotypes for all unobserved markers in a given ancestry tree without assumptions on the structure of the tree. Linkage modeling is localized to each imputed marker, as opposed to creating a global model of linkage disequilibrium or haplotypes for each chromosome. Multi-generation imputation is performed by representing the pedigree tree as a modified Bayesian network. A graph traversal heuristic allows the approach to scale to arbitrarily large ancestries. Our internal implementation utilizes several modern technologies to perform efficiently at scale, including Neo4j and Apache Spark. The accuracy of our technique is shown to outperform Beagle, an industry-standard imputation method, in many realistic scenarios.