Improved assembly of noisy long reads by k-mer validation

Genome assembly depends critically on read length. Two recent technologies, PacBio and Oxford Nanopore, produce read lengths above 20 kb, which yield genome assemblies that are vastly superior to those based on Sanger or short-reads. However, the very high error rates of both technologies (around 15%-20%) makes assembly computationally expensive and imprecise at repeats longer than the read length. Here we show that the efficiency and quality of the assembly of these noisy reads can be significantly improved at a minimal cost, by leveraging on the low error rate and low cost of Illumina short reads. Namely, k-mers from the PacBio raw reads that are not present in the Illumina reads (which account for ~95% of the distinct k-mers) are deemed as sequencing errors and ignored at the seed alignment step. By focusing on ~5% of the k-mers which are error-free, read overlap sensitivity is dramatically increased. Equally important, the validation procedure can be extended to exclude repetitive k-mers, which avoids read miscorrection at repeats and further improve the resulting assemblies. We tested the k-mer validation procedure in one long-read technology (PacBio) and one assembler (MHAP/ Celera Assembler), but is likely to yield analogous improvements with alternative long-read technologies and overlappers, such as Oxford Nanopore and BLASR/DAligner.