Expecting loss but getting CNLoH

[tinakeshav]

Hi,
I've managed to use your package successfully with some suggested tweaks in #74; thank you very much!
I'm getting CNLoH for a very commonly deleted region in my samples and the deletion is confirmed with WGS -- recurrent across independent samples from different patients.
Some samples show deletion in that region in specific clones. But in many, multiple clones only present with CNLoH in that region and I'm fairly certain there's a deletion there.
Why might this be the case? Are there any specific parameters that require tuning?
Thank you so much!!

[teng-gao]

Hello,

The CNV type assignment of CNLoH vs Deletion is decided by the expression change (logFC). So a better diploid expression reference that more closely matches the cell types / sequencing platforms of your target samples can help with this.

[tinakeshav]

Hi Teng, Thank you for your suggestion. Using a different reference marginally helped with this but for many samples it remains a big problem. exp_roll_clust.png seems to agree with the expected results quite well.
Is there a way to lower the probability of CNLoH detection / other parameters that I can tune to detect the losses/amplifications that are being called as CNLoH?

[teng-gao]

Hi @tinakeshav ,

Could you show some examples (e.g. HMM bulk profiles)? The logphi_min threshold in the analyze_bulk subroutine (for performing CNV calling on pseudobulks) determines the minimum deviation in expression magnitude for calling DEL/AMP instead of CNLoH:
https://kharchenkolab.github.io/numbat/reference/analyze_bulk.html

However this threshold is usually fixed in the main workflow so you'd probably have to hack it in (e.g. passing the argument from run_numbat).

[teng-gao]

Hi @tinakeshav ,

Starting from Numbat v1.3.0, you can specify WGS-derived CNV calls via segs_consensus_fix. The CNV states and boundaries for the segments you provide will stay fixed. Let me know how it goes if you end up giving it a try!

[tinakeshav]

Hi Teng,
Thank you so much! I will certainly give this a go and share the results with you. A bit of a naive question, but is there a convention for supplying the segment IDs? (Is there maybe a genomic interval --> segment ID dictionary somewhere?) Looking forward to trying it

[teng-gao]

You can check this function - the segments are named as chromsome + letter postfix, e.g. 1a, 1b, etc. Following this naming convention isn't required for the provided segments though.