I have got a quite large VCF file (about 20k samples with 160k SNPs after filtering for quality etc.) and I would like to get a phylogeny for it. However, the whole dataset's too large for my computational resources (unless I go for fasttree which would be my last resort). Therefore, I need to subset the SNPs and there's plenty of possible ways to do this (dump all rare SNPs with MAF below a threshold, only use SNPs in coding regions, only use missense SNPs; just to name a few).

Is there a consensus (or at least some comparative work or recent review; I found nothing) on what method works best and gives the tree that's closest to using the full dataset?

  • 1
    $\begingroup$ Yes, the SNPs have already been filtered on a handful of metrics. $\endgroup$
    – zeawoas
    Jan 9 at 15:05
  • 2
    $\begingroup$ OK - what species are you working on? It's worth looking at a paper that tries to build a tree using the same species as you and then following some of the steps they have performed. For example this paper if you are working with viruses. $\endgroup$
    – user438383
    Jan 9 at 17:05

I will give an answer that is specific to closely-related microbial genomes. Core-SNPs, SNPs which are within conserved and likely orthologous sequences, have been shown to be relatively informative for constructing phylogenies of large sequence sets[1].

Parsnp is the most popular tool for constructing core-genome alignments on microbial genomes, and extracts the core SNPs from the alignments. Parsnp then uses raxml or FastTree in order to produce a phylogeny from the core SNPs. I will say, though, that 20k sequences may be too many for Parsnp to run on (depending on how large the sequences are and your available resources).

Disclaimer: I am the maintainer of Parsnp (:


The subset that gives you the tree closest to the one you obtain with the full dataset does not necessarily is the tree nearest to reality... It could be that you get a more realistic result using only partial data for multiple reasons. Your best chance is in my opinion, as it usually is, making lots of tests and contrasting hypotheses.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.