HLA Allele and LoH Calling on WXS Tumors

[andrewgalbraith21]

Hello,

I am hoping to run HLA allele and LoH calling on ~100 tissue and blood matched WXS tumor samples. Similar, to #47 , I am having an issue on around 10 samples where there are too many blocks and possible haplotypes after phasing. I can get these sample to run by increasing this the threshold up from 512 in the source code. However, this quick fix does not always work as a few of my more mutation high tumor samples result in over 10,000,000 possible haps which makes them very time and memory intensive to process. As such I have few questions to ask which I think might help to work around this issue.

1. Are there any preprocessing steps which may lower the number of blocks before running SpecHLA.sh? I do use ExtractHLAread.sh on the bam file before running

2. Are there any parameters adjustments for SpecHLA.sh which may help lower the number of HLA blocks while retaining accurate HLA allele calls? Currently, I have altered some like -s but have had limited success

3. Is there any consideration to alter the approach for detecting HLA LoH in tumor samples. Right now, my understanding is that you need to run SpecHLA.sh on both the normal and tumor sample and then input the normal hla calls along with the tumor hla frequency files into cal.hla.copy.pl. However, I think a more practical method would be to only perform SpecHLA.sh on normal samples and then have a script which realigns tumor reads to the normal hla references and gathers relative allele frequencies. This approach is more akin to LOHHLA and is less likely to have problems when running on tumor samples wherein 1) hla reconstruction is very difficult or computationally intensive in the tumor and 2) hla types are discordant between tumor and normal.

Please respond to this issue with your thoughts and answers when you have the chance thanks!

[wshuai294]

Hi, to reduce the number of blocks, you can set a lower value of -m and set a higher value of -r and -q. Also, filter the sequencing reads before running SpecHLA using fastp might help. I am wondering, in the over 10,000,000 possible haps case, which gene is it? How many heterozygous variants for this? I guess there might be a lot false positive variants. For the LOH issue, your suggestion is very reasonable. But it requires additional effort, as aligning tumor reads to two normal cell alleles might lead to ambiguity; some reads can be mapped to two alleles with similar identity. Some issues like this should be addressed.