Assembly of Large Genomes from Paired Short Reads

The de novo assembly of genomes from high-throughput short reads is an active area of research. Several promising methods have been recently developed, with applicability largely restricted to the smaller and less complex bacterial genomes. In this paper, we present a method for assembling large genomes from high-coverage paired short reads. Our method exploits large distributed memory and parallelism available on multiprocessor systems to handle memory-intensive phases of the algorithm, effectively allowing scaling to large genomes. We present parallel algorithms to construct a bidirected string graph that is several orders of magnitude smaller than the raw sequence data and to extract features from paired reads. We also present a heuristic method that uses these features to guide the extension of partial graph traversals corresponding to large genomic contigs. In addition, we propose a simple model for error correction and derive a lower bound on the coverage needed for its use. We present a validation of our framework with short reads from D. melanogaster and S. cervisiae synthetically generated at 300-fold coverage. Assembly of the D. melanogaster genome resulted in large contigs (50% of the genome covered by contigs larger than 102Kb), accurate to 99.9% of the bases, in under 4 hours of wall clock time on a 512-node Blue Gene/L.