SeqWare Query Engine: storing and searching sequence data in the cloud

Abstract

Background: Since the introduction of next-generation DNA sequencers the rapid increase in sequencer throughput, and associated drop in costs, has resulted in more than a dozen human genomes being resequenced over the last few years. These efforts are merely a prelude for a future in which genome resequencing will be commonplace for both biomedical research and clinical applications. The dramatic increase in sequencer output strains all facets of computational infrastructure, especially databases and query interfaces. The advent of cloud computing, and a variety of powerful tools designed to process petascale datasets, provide a compelling solution to these ever increasing demands.

Results: In this work, we present the SeqWare Query Engine which has been created using modern cloud computing technologies and designed to support databasing information from thousands of genomes. Our backend implementation was built using the highly scalable, NoSQL HBase database from the Hadoop project. We also created a web-based frontend that provides both a programmatic and interactive query interface and integrates with widely used genome browsers and tools. Using the query engine, users can load and query variants (SNVs, indels, translocations, etc) with a rich level of annotations including coverage and functional consequences. As a proof of concept we loaded several whole genome datasets including the U87MG cell line. We also used a glioblastoma multiforme tumor/normal pair to both profile performance and provide an example of using the Hadoop MapReduce framework within the query engine. This software is open source and freely available from the SeqWare project (http://seqware.sourceforge.net).

Conclusions: The SeqWare Query Engine provided an easy way to make the U87MG genome accessible to programmers and non-programmers alike. This enabled a faster and more open exploration of results, quicker tuning of parameters for heuristic variant calling filters, and a common data interface to simplify development of analytical tools. The range of data types supported, the ease of querying and integrating with existing tools, and the robust scalability of the underlying cloud-based technologies make SeqWare Query Engine a nature fit for storing and searching ever-growing genome sequence datasets.

Figure 1

SeqWare Query Engine schema. The HBase database is a generic key-value, column oriented database that pairs well with the inherent sparse matrix nature of variant annotations. (a) The primary table stores multiple genomes worth of generic features, variants, coverages, and variant consequences using genomic location within a particular reference genome as the key. Each genome is represented by a particular column family label (such as “variant:genome7”). For locations with more than one called variant the HBase timestamp is used to distinguish each. (b) Secondary indexing is accomplished using a secondary table per genome indexed. The key is the tag being indexed plus the ID of the object of interest, the value is the row key for the original table. This makes lookup by secondary indexes, “tags” for example, possible without having to iterate over all contents of the primary table.

Figure 2

Load and query performance. Comparisons of load and query times between the HBase and BerkeleyDB backend. (a) Load times for the “1102 GBM” tumor/normal genomes where compared between HBase and BerkeleyDB. Both used a single-threaded approach to better compare relative performance. Both perform similarly but over time the load times for BerkeleyDB increase faster than with HBase. (b) Comparison of querying the 1102 genome database between BerkeleyDB, HBase single threaded, and HBase using MapReduce. Beyond 3M variants BerkeleyDB query times increase dramatically while both query types for HBase perform linearly, with MapReduce consistently exhibiting the best performance.