Abstract
The recently developed next generation sequencing platforms not only decrease the cost for metagenomics data analysis, but also greatly enlarge the size of metagenomic sequence datasets. A common bottleneck of available assemblers is that the trade-off between the noise of the resulting contigs and the gain in sequence length for better annotation has not been attended enough for large-scale sequencing projects, especially for the datasets with low coverage and a large number of nonoverlapping contigs. To address this limitation and promote both accuracy and efficiency, we develop a novel metagenomic sequence assembly framework, DIME, by taking the DIvide, conquer, and MErge strategies. In addition, we give two MapReduce implementations of DIME, DIME-cap3 and DIME-genovo, on Apache Hadoop platform. For a systematic comparison of the performance of the assembly tasks, we tested DIME and five other popular short read assembly programs, Cap3, Genovo, MetaVelvet, SOAPdenovo, and SPAdes on four synthetic and three real metagenomic sequence datasets with various reads from fifty thousand to a couple million in size. The experimental results demonstrate that our method not only partitions the sequence reads with an extremely high accuracy, but also reconstructs more bases, generates higher quality assembled consensus, and yields higher assembly scores, including corrected N50 and BLAST-score-per-base, than other tools with a nearly theoretical speed-up. Results indicate that DIME offers great improvement in assembly across a range of sequence abundances and thus is robust to decreasing coverage.