Abstractions for Genomics: Or, which way to the Genomic Information Age?

With DNA full sequencing costs dropping below $1, 000, two major application drivers are: finding the genetic basis for disease (via Genome Wide Association Studies), and personalized medicine (tailoring treatments to genetic makeup as in cancer genomics). While hardware costs are falling, software costs are a major obstacle to both discovery and personalized medicine. In this paper, we propose attacking this obstacle via the use of a hierarchical set of abstractions (akin to Internet layers) for genome processing. While interface formats exist for mapped files (BAM) and variant formats (VCF), we propose a new separation of variant calling into a deterministic evidence layer that is queried by a probabilistic inference layer via a proposed Genome Query Language. GQL allows inference engines to efficiently sift only the relevant evidence (e.g., the READs that flank a SNP site) from raw genomic data. We describe a set of exemplar queries ranging from SNP calling to phasing based on input from sequencing compa-nies and show how evidence for these queries can be concisely captured using GQL. Discovery is modeled as finding the variants common to a given phenotype, and personalized medicine as the dual problem of finding the phenotypes corre-sponding to individual variations, from a database of complete individual genome sequences. We suggest that the standard implementation of GQL using relational database tables with billions of rows for each genomic location will not scale to a single human; instead, we introduce efficient indices based on compressed strength vectors to scale to large populations. W