Suggested model with no individual-level training data

[Fiwx]

Hi Florian,

What is the best LDpred2 model for generating a PRS using only sumstats and a reference panel (i.e., no individual-level training/validation data)? I'd like to construct a PRS without any individual-level data, and then test it using individual-level data; I assume LDpred2-auto should be of use. Also, would you expect this method to still have decent accuracy if already thresholded summary statistics (500-20,000 SNPs) are being used? Thank you!

[privefl]

Yes, you should use LDpred2-auto. I don't know whether it is a good idea to apply it on already thresholded data. But, if you try it, you should probably use snp_thr_correct(), as I did in https://doi.org/10.1016/j.ajhg.2021.11.008.

[privefl]

Any update on this?

[Fiwx]

Thank you! That is very helpful information.