Optimizing k-mer size using a variant grid search to enhance de novo genome assembly

Largely driven by huge reductions in per-base costs, sequencing nucleic acids has become a near-ubiquitous technique in laboratories performing biological and biomedical research. Most of the effort goes to re-sequencing, but assembly of de novogenerated, raw sequence reads into contigs that span as much of the genome as possible is central to many projects. Although truly complete coverage is not realistically attainable, maximizing the amount of sequence that can be correctly assembled into contigs contributes to coverage. Here we compare three commonly used assembly algorithms (ABySS, Velvet and SOAPdenovo2), and show that empirical optimization of k-mer values has a disproportionate influence on de novo assembly of a eukaryotic genome, the nematode parasite Meloidogynechitwoodi. Each assembler was challenged with about 40 million Iluumina II paired-end reads, and assemblies performed under a range of k-mer sizes. In each instance, the optimal k-mer was 127, although based on N50 values,ABySS was more efficient than the others. That the assembly was not spurious was established using the "Core Eukaryotic Gene Mapping Approach", which indicated that 98.79% of the M. chitwoodi genome was accounted for by the assembly. Subsequent gene finding and annotation are consistent with this and suggest that k-mer optimization contributes to the robustness of assembly.