kmer vs accuracy

[Autotroph]

Hi,

Given below is an explanation that i found for the different kmer size and its effect on sensitivity and specificity.

"A smaller k-mer value allows locating more overlapping
sequence (i.e. higher sensitivity) while increasing the number of ambiguous repeats(i.e. lower specificity). An extreme example for that is choosing a k-mer length of 3 or 5 where there will be only few possible high coverage k nucleotides with very low specificity. On the other hand, choosing a higher k-mer value has the opposite effect - it decreases sensitivity while increasing specificity.”

From this i understand that an assembly at a higher kmer size is always more "accurate"(not talking about better N50) than the one at a lower kmer size. Is that correct(irrespective of the sequencing errors and polymorphism)?

Could anybody explain how the accuracy of an assembly goes down with lower kmer size?

Thanks

[Thorondor]

yes that is mostly correct.

just check out how brujin graph works, it's not hard to understand. Daniel Zerbinos phd thesis is helpful: https://www.ebi.ac.uk/training/ftp/P...el_Zerbino.pdf

the larger the kmer the longer the overlap between two reads has to be. that's also a reason why the kmer can never be larger then your minimum read length.

If you have a high coverage you can choose a higher kmer which will result in more reliable and accurate contigs.

So with kmer size 6 you will have nodes like: AAGCTG, GGCTTA, CTCAAG... and of course there will be more overlaps (nodes need to ovlerap with kmer-1 bps to be contected in most algorithms) then with kmer size 35: AGTTGGCTAGAAACTGTACTAGTTTCGCGCATGCA and you will have less nodes therefore less RAM is needed.

[Autotroph]

Thanks Thorondor.