Accelerating De Novo Assembler WTDBG2 on Commodity Servers

Abstract

De novo genome assembly reconstructs the chromosomes from massive relatively short fragmented reads and serves as fundamental for studying new species where there is no reference genome. Wtdbg2 is a de novo assembler for long reads that is up to hundreds of kilobases. It is based on fuzzy-Bruijn graph (FBG) and is ten times faster than the cutting-edge assemblers such as Canu. However, the performance of wtdbg2 still requires further improvement: 1) it requires up to terabytes of memory to compute the assembly, which is infeasible to run on commodity server; 2) it requires tens of hours for assembling on large datasets such as genomes of homo sapiens. To address the above drawbacks, we propose several optimization techniques for accelerating wtdbg2 on commodity server, including a memory auto-tuning scheme, sequence alignment optimization and intermediate result elimination in the output procedure. We compare the optimized wtdbg2 with the original implementation and two cutting-edge assemblers on real-world datasets. The experiment results demonstrate that optimized wtdbg2 achieves maximum and average speedup of 2.31$\times$ and 1.54$\times$ respectively. In addition, our proposed optimization reduces the memory usage of wtdbg2 by 39.5% without affecting the correctness.