Question about some operations

[shy218]

What is the math formula for PC-Relate in relatedness estimation?

Is the LD prune supported?

[hammer]

@ravwojdyla knows the PC-Relate code the best.

LD-prune is being actively worked on right now by @tomwhite. He’s using the original implementation by @eric-czech as a template.

[shy218]

Sorry for any confusion. I also have a question for variant normalization. Does variant normalization mean to make sure the variant is parsimonious and left-aligned? Is there any implementation in sgkit? Sorry if I ask a stupid question.

[hammer]

Yes, that's roughly what variant normalization means. It might also mean "lifting" the variant onto another version of the reference genome, or a few other operations. We do not have any of those operations in sgkit currently.

[eric-czech]

For HWE, are you using the equation from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1199378/?

Hey @shy218, a few more details on that:

The calculation is integrating over an exact distribution that is described pretty succinctly at https://hail.is/docs/0.2/_static/LeveneHaldane.pdf. I ported the implementation in the paper you mentioned from the C version at http://csg.sph.umich.edu/abecasis/Exact/index.html. I doubt the specifics will interest you much but I am certainly curious how this one performs since I haven't compared this to Hail/PLINK (I have a vague idea how several other operations perform).

I missed this on the call but HWE performance depends on the data, much like LD-prune -- I forgot to mention that. The others we discussed are only dependent on array shapes. It is easy to generate data for realistic or worst-case scenarios though.

[shy218]

I write the algorithm by steps in Math for HWE. Do you think this understanding is correct? And also, do you think it is a good idea to test on HWE compared with Hail/PLINK?

[eric-czech]

Do you think this understanding is correct?

Not sure, you'd have to check that against the paper or the code.

And also, do you think it is a good idea to test on HWE compared with Hail/PLINK?

By test do you mean for accuracy or for speed? We already do the former and yes the latter would be useful.

[ravwojdyla]

👋 @shy218 see the equation 3 and 4 in https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4716688/

[shy218]

Thanks for the reply. I understand the formula now.