Suggestions for WGS 5X Sequences

dilawerkh4dilawerkh4 Member
edited July 2017 in Ask the GATK team

Hi Geraldine or Sheila,

I am in the process of customizing a GATK pipeline for processing aDNAA. I have processed a couple of 3000 year old WGS sequences so far using GATK best practices, and although the resulting VCFs have been ok, they are far from perfect, and I was looking for any suggestions you can offer on how to minimize reference bias both during alignment as well as variant calling. To give you an idea where things stand, I have organized a color coded spreadsheet summarizing CollectVariantCallingMetrics using various GATK tools. Here is a link to the sheet https://docs.google.com/spreadsheets/d/1iPw-afPV6Z4zGzqfOlna_UyYf87gBv1ghxzLzDkqfFw/edit?usp=sharing

The attached spreadsheet will show you the metrics with and without BQSR, and with or without VQSR.

I have also experimented with omitting some of the annotations you use in VQSR and have plotted everything in R, although the output with all annotations using -mG 4 was not bad.

So far my outputs have been using HaplotypeCaller in GVCF and joint genotyping using 2 or 3 samples, I plan to increase to 10 to see if that helps establish better evidence for variants at low depth sites

Any suggestions you can offer that I may try to reduce reference bias at low depth positions. Filtering out positions that passed VQSR but had a QUAL score <100 with vcftools helped some. Any suggestions with regards changing the prior likelihood of true and not true training sites and known sites during VQSR or to use different resources than the HapMap 1000G or Omni that you suggest.

Post edited by dilawerkh4 on
Tagged:

Best Answer

Answers

  • SheilaSheila Broad InstituteMember, Broadie, Moderator

    @dilawerkh4
    Hi,

    Are you saying you are seeing too many reference calls in your dataset that you expect to be variant calls? There are a few parameters you can play around with in HaplotypeCaller, such as --heterozygosity. Adding in the extra samples will also help to make more confident variant calls.

    I am unclear why you are filtering out sites which have QUAL < 100 if you are expecting more variant calls? As stated in the other thread, we do not have much experience with aDNA, so you will have to do some experimenting on your own. You may also try manual review in IGV.

    -Sheila

  • dilawerkh4dilawerkh4 Member
    edited August 2017

    @Sheila said:
    @dilawerkh4
    Hi,

    Are you saying you are seeing too many reference calls in your dataset that you expect to be variant calls? There are a few parameters you can play around with in HaplotypeCaller, such as --heterozygosity. Adding in the extra samples will also help to make more confident variant calls.

    I am unclear why you are filtering out sites which have QUAL < 100 if you are expecting more variant calls? As stated in the other thread, we do not have much experience with aDNA, so you will have to do some experimenting on your own. You may also try manual review in IGV.

    -Sheila

    Hi Sheila,

    Thanks much for pointing out the --hetrozygosity parameter. Please correct me if I am wrong, but my understanding of the parameter is that if we increase it from its default of 0.001 ( two randomly chosen chromosomes from the population of organisms would differ from each other at a rate of 1 in 1000 bp) to say 0.01 (rate of 1 in 100 bp) HaplotypeCaller will allow for more mismatched alleles when mapping to reference (ie a site will be called if it is non-reference), which in effect should decrease reference bias.

    With regards to filtering QUAL<100, I agree it is too aggressive, I will lower it to say 25, and see what happens, but the reasoning behind it was aDNA is riddled with artifacts, as well as substitutions at the read ends of C -->T and G -->A due to post mortem deamination, which I deal with by checking sites for which the reference is C or G and the sample is T or A. Although, I try to downgrade quality scores at those sites using MapDamage, there are some that are not filtered. Regardsless, I agree that filtering for QUAL <100 is too aggressive.

    Also, there appears to be a Picard tool, AlignmentSummaryMetrics, that is not talked about here, that may help with these types of stats, what do you think?

    Dilawer

Sign In or Register to comment.