Quality control of summary statistics

[JingZhang1227]

Hi Florian,

Sorry for another question.

I am using ldpred2 to calculate polygenic risk scores. I am running the automatic model with the provided LD reference based on the example-with-provided-ldref.R.

I have three datasets: the summary statistics, the provided LD reference, and a test dataset (plink format). The positions of both the summary statistics and the test dataset are on hg38.

Based on the example-with-provided-ldref.R line 27, I will perform the quality control of summary statistics using the allele frequency from the provided LD reference with is “af_UKBB” to get sd_ldref. Since my test dataset is different from the provided LD reference, I just want to double check if I should also calculate and use standard deviations of genotypes of individuals in the test dataset in this step. My test dataset is family data. Individuals within a family are related (father, mother, and children).

Thank you very much!

[privefl]

I guess you can use both.
Checking with af_UKBB will check if you have some consistency between sumstats and LD ref, and checking with sd in your test data will check for this consistency between sumstats and test set.

[JingZhang1227]

Thanks for getting back to me!

I saw in your prepare-sumstats.R line 16, you used sqrt(big_colstats(G, ind.val, ncores = NCORES)$var) to calculate the standard deviations of genotypes of individuals in the validation dataset. I have related individuals in my test dataset, I was wondering if I could use the same command.

Thanks again!

[privefl]

Yeah I guess related individuals might reduce the standard deviation.
You can probably use sqrt(2 * af * (1-af)) instead, as i don't think the allele frequency is biased by relatedness.

[JingZhang1227]

Thank you very much! This is very helpful! Sorry I think I have just one more related question. For the allele frequency, or af, I think I can get it using big_colstats(G, ncores = NCORES)$sum / (2 * nrow(G)) (from the provide-ld-ref.R line 15). Or I was wondering if you would suggest me calculate af using something like plink --freq function where the minor allele frequencies is calculated only based on founders (individuals without parents) in the test sample.

[privefl]

Yeah you can use PLINK if you prefer.
Call it directly from R using system(glue::glue()) :')

[privefl]

Any update on this?

[JingZhang1227]

Sorry for the late response. Thank you very much for your suggestions! I used the PLINK MAF! I had another unrelated question so I submitted a new issue which is issue #200.

Thank you very much again!