Towards accurate de novo assembly for genomes with repeats

De novo genome assemblers designed for short k-mer length or using short raw reads are unlikely to recover complex features of the underlying genome, such as repeats hundreds of bases long. We implement a stochastic machine-learning method which obtains accurate assemblies with repeats and self-validates assemblies via consensus. For this, a prior assembler is extended with the ability to (a) assemble variable-length raw reads, which may span and unambiguously recover interspersed repeats in the genome, and (b) recognize long, direct terminal repeats during the assembly, then report an unambiguous circular assembly. Consensus is obtained via stochastically independent runs of the assembler on the same read library. We experiment on viral and mitochondrial genomes of up to 41 kbp, with synthetic raw-read libraries, to be able to evaluate the assembly against a reference. We show the prerequisites for obtaining accurate assemblies. For genomes with interspersed repeats, using raw reads of average length comparable to the repeat length likely gives an accurate genome. Genomes with long direct terminal repeats can be assembled accurately also with reads shorter than the repeat length. In both cases, a simple majority forms consensus, since over 70 % of independent runs on this set of genomes yield a correct assembly.