Should Genome Assembly Improve with More and More Reads?

Conventional wisdom says yes, but that may not work for de Bruijn graph-based assemblers. Last year, we were experimenting with Velvet and noticed something strange. When we randomly removed half of the reads from a genomic library, the ‘assembly quality’ seemed to have improved. Assembly quality was measured in terms of N50 of scaffolds and not using any other detailed analysis. What was going on?

We did k-mer analysis of the assembled scaffolds. It was clear that the assembly was wrong, because some k-mers supposedly present only once in the genome were found more than 5-6 times in the assembly. Here was our best explanation for the observation. A de Bruijn graph-based assembler splits all reads into k-mers and forms a giant graph. That graph is resolved into contigs. At times, the program sees overlaps between two branches (such as the second figure here) and creates multiple contigs all terminating at the junction in the figure.

What happens, when you remove some reads? The data becomes sparse and one or other branch of the graph may not appear due to sparseness. The assembler does not pause at the junction and evaluates one or other branch as a contiguous sequence, even when it is not.

The second best explanation was a possible defect in Velvet’s scaffolding routine. We did not have to explore further to resolve between the possibilities.