Efficiency and Power as a Function of Sequence Coverage, SNP Array Density, and Imputation

High coverage whole genome sequencing provides near complete information about genetic variation. However, other technologies can be more efficient in some settings by (a) reducing redundant coverage within samples and (b) exploiting patterns of genetic variation across samples. To characterize as many samples as possible, many genetic studies therefore employ lower coverage sequencing or SNP array genotyping coupled to statistical imputation. To compare these approaches individually and in conjunction, we developed a statistical framework to estimate genotypes jointly from sequence reads, array intensities, and imputation. In European samples, we find similar sensitivity (89%) and specificity (99.6%) from imputation with either 1× sequencing or 1 M SNP arrays. Sensitivity is increased, particularly for low-frequency polymorphisms (), when low coverage sequence reads are added to dense genome-wide SNP arrays — the converse, however, is not true. At sites where sequence reads and array intensities produce different sample genotypes, joint analysis reduces genotype errors and identifies novel error modes. Our joint framework informs the use of next-generation sequencing in genome wide association studies and supports development of improved methods for genotype calling. Author Summary: In this work we address a series of questions prompted by the rise of next-generation sequencing as a data collection strategy for genetic studies. How does low coverage sequencing compare to traditional microarray based genotyping? Do studies increase sensitivity by collecting both sequencing and array data? What can we learn about technology error modes based on analysis of SNPs for which sequence and array data disagree? To answer these questions, we developed a statistical framework to estimate genotypes from sequence reads, array intensities, and imputation. Through experiments with intensity and read data from the Hapmap and 1000 Genomes (1000 G) Projects, we show that 1 M SNP arrays used for genome wide association studies perform similarly to 1× sequencing. We find that adding low coverage sequence reads to dense array data significantly increases rare variant sensitivity, but adding dense array data to low coverage sequencing has only a small impact. Finally, we describe an improved SNP calling algorithm used in the 1000 G project, inspired by a novel next-generation sequencing error mode identified through analysis of disputed SNPs. These results inform the use of next-generation sequencing in genetic studies and model an approach to further improve genotype calling methods.