13
$\begingroup$

I have estimated genome-wide heterozygosity levels using maximum likelihood and classical substitution model implemented in package atlas. These estimates are way more robust than classical SNP calling and simply counting number of heterogeneous locations because they integrate information from a big window into one estimate. I.e. I am quite sure about heterozygosity levels of SNPs of my genomes.

However, now I would like to know also about heterozygosity of indels and call individual snps using my robust estimates as a prior. GATK allows to specify priors on heterozygosity both of SNPs and indels, which might improve variant calls a lot, because default GATK is parametrised for human (Default priors for heterozygous SNPs is 0.001 and for indel 1.25E-4.). But, I work with insects I know that my SNPs heterozygosity is somewhere between 0.01 and 0.03 (1% - 3%).

So here comes the question, if I know the heterozygosity of SNPs, can I simply guess a reasonable prior for the heterozygosity of indels? GATK defaults for priors are 10x different, should I assume the same relation? Is there a reasonable way how to guess a prior for heterozygosity of indels?

Improve this question
$\endgroup$
0

1 Answer 1

Reset to default
1
$\begingroup$

SNPs, insertions, and deletions are generated by different biological processes (see for example here). I would expect the per-generation rates of these mutational processes to vary across species. I would also expect the degree to which each mutational process varies across species to vary as well. I think the answer to your question is "probably not without further information".

However, priors are a feature of Bayesian statistics. In Bayesian inference, it is good practice to redo one's analysis using several priors to establish how sensitive one's results are to the chosen prior. I would suggest that you try to call indels using a range of priors and compare the quality of your call sets to determine how important the choice of prior actually is.

Improve this answer
$\endgroup$
5
  • $\begingroup$ 1, The processes that generate mutations are indeed different, but the processes maintaining them are the same (drift, selection). 2, Looking at result by different prior is sort of equivalent to using maximum likelihood, no? What would be the point of priors if we had no way how to specify them meaningfully? $\endgroup$
    – Kamil S Jaron
    Commented Nov 17, 2018 at 9:45
  • $\begingroup$ I am starting to think, that it's just not a good question because it does not have a clear "yes/no" answer. $\endgroup$
    – Kamil S Jaron
    Commented Nov 17, 2018 at 9:45
  • $\begingroup$ Maybe it's worth rephrasing the question to "how do I..." or "What information do I need in order to..." $\endgroup$
    – winni2k
    Commented Nov 17, 2018 at 20:05
  • $\begingroup$ I thought that posterior probability is likelihood times prior. Then trying multiple priors (or taking a non-informative one) is kind of the same thing as just calculating maximum likelihood estimate without the prior. Or have I missed something? $\endgroup$
    – Kamil S Jaron
    Commented Nov 18, 2018 at 1:25
  • $\begingroup$ I see. ML (together with the bootstrap!) and Bayesian inference are both statistical methods that are commonly used to solve similar problems. In that regard, they are "kind of the same thing". As to the second question, "What would be the point of priors if we had no way how to specify them meaningfully?": Of course, there would be none ;) But the point is that sometimes we can, and the Bayesian approach works regardless of whether you have a strong prior or not. $\endgroup$
    – winni2k
    Commented Nov 18, 2018 at 8:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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