Difference between Polysolver and mhcflow

[ysbioinfo]

Hi @svm-zhang ,
I tested MHCflow on some WES samples (150X, aligned to hg19) and noticed that the results differ from those generated by Polysolver. Specifically, MHCflow seems to favor calling homozygous loci more often. I pasted the result from one sample below as an example:
mhcflow result:

allele  gene    tot_scores      sample
hla_a_30_01_01  hla_a   8022770.0758    P022_005
hla_a_30_01_01  hla_a   4011385.0379    P022_005
hla_b_40_02_11  hla_b   939190.0302     P022_005
hla_b_40_02_15  hla_b   1813238.3457   P022_005
hla_c_07_02_02  hla_c   191286.0893     P022_005
hla_c_07_17_02  hla_c   326942.8309     P022_005

polysolver result:

HLA-A   hla_a_11_01_01  hla_a_30_01_01
HLA-B   hla_b_40_02_01  hla_b_13_02_01
HLA-C   hla_c_03_04_01_01       hla_c_06_02_01_01

The FASTQ file and kme files were retrieved from Polysolver data, and I manually created the HLA class I BED file as follows

chr6	29909037	29913661
chr6	31321649	31324965
chr6	31236526	31239907

I’m aware you've made improvements upon Polysolver, so I’m wondering: is this kind of discrepancy expected? I would be happy to provide any additional files or information if that would help.

I also have another question, which may be outside the scope of mhcflow, but I would greatly appreciate any insights you could offer:

I previously used OptiType for HLA typing. However, since OptiType only provides 4-digit HLA types, and the original LOHHLA script requires 8-digit HLA types (because the corresponding FASTA files are categorized by 8-digit alleles), what would be the best way to adapt OptiType results for LOHHLA? I’ve seen some approaches that replace the 4-digit allele with the longest 8-digit form, does this method make sense from your perspective?

Thanks again for your amazing work and for any advice you can share!

Best regards,
Yang

[svm-zhang]

Hello Yang,

Thanks for reporting.

For the observed discrepancy, I can suggest following things to check:

• Using samtools flagstat to check if the realigned BAM files from mhcflow and polysolver have the same alignment stats.
• If they do, please check the score table generated by mhctyper in the typer folder with the score files produced by polysolver. For example, what is the score of hla_a_11_01_01 in the mhctyper table? For the different alleles called by polysolver, are they among the top-5 alleles in the mhctyper score table?
• There is a v0.1.0 branch in this repo, you can clone and try that version on your data. The v0.1.0 version should have the same result with polysolver, at least for the 1000g samples I tested. If the v0.1.0 version generates the same result as the one produced by polysovler, I need to test mhcflow on more samples on my end (which can take longer time).

By the end of the day, I cannot fully debug without the data.

I will get back to your other questions tomorrow. Thanks for your patience.

Simo

[ysbioinfo]

Thanks Simo, for your prompt response. I'll check those files according to your suggestions after the holiday.

Yang

[svm-zhang]

I have another note for the observed discrepancy: what is the value you passed to the —min-ecnt option? By default, all realignments contribute to the typing process, which is not exactly the same strategy used in the polysolver. If you have not used this option, could you try to rerun mhcflow using —min-ecnt 1?

This should be the second thing to check once you confirm the realigned BAM files have the same alignment records using samtools flagstat.

[svm-zhang]

I am moving the discussion on Optitype to another ticket and let us continue the discuss over there.