Running LDpred2 auto with reference sample

privefl / bigsnpr

R package for the analysis of massive SNP arrays.

https://privefl.github.io/bigsnpr/

186 stars 44 forks source link

Running LDpred2 auto with reference sample #413

Closed lm380-r closed 1 year ago

lm380-r commented 1 year ago

Dear Dr. Prive,

Would you be able to share an example of running LDpred2 with the newer LD reference panel you provide on your figshare with the PGC schizophrenia summary statistics?

I have no issues calculating the PRS when I use our in-house data to create the correlation but for some reason, the PRS are massively inflated when I try to create a PRS with the LD reference panel. I have tried various summary statistics and they all have the same issue. I follow the older code you provide (example-with-provided-ldref.R) so I'm not sure what is going on.

Thank you in advance!

privefl commented 1 year ago

What do you mean by "massively inflated"?

lm380-r commented 1 year ago

The PRS values range from x10^-17 to x10^17, which seems peculiar for PRS and there is no correlation between the PRS developed with the ld reference panel and those created with only our in-house data.

privefl commented 1 year ago

Okay, this has been discussed in other issues here, but I'm repeating it here.

You need to:

update to the latest GitHub version of the package -> you will quickly get NAs instead of these very large values
make sure to use the robustness parameters (allow_jump_sign = FALSE, shrink_corr = 0.95) -> you can even decrease shrink_corr to e.g. 0.5 if there is a strong mismatch between the LD and GWAS sumstats
if the above still does not help, you can additionally use use_MLE = FALSE to revert back to some previous sampling strategy that is less flexible but more robust

privefl commented 1 year ago

Any update on this?