LDpred2 cross-sample overlap?

[emreardadeniz]

Dear @privefl

I have a sample overlap problem between the base data (summary stats) and the target data (ALSPAC UK) when generating polygenic risk scores using PRSice. Does LDpred2 control for this sample overlap? Somewhere I read LDpred2 applies a Bayesian approach to adjust effect sizes based on LD structure, which somehow mitigates, although not fully, the effects of overfitting caused by base-target overlap? What are your opinions and do you think LDpred2 is what we need here?

Thanks
Emre

[privefl]

So, basically, there are individuals in your test set who were also part of the GWAS you're using to make the PGS?
No PGS method can correct for this I think; this will always cause overfitting.
But, there might be some methods trying to residualize GWAS summary statistics for another set. Have a look e.g. at

• https://pmc.ncbi.nlm.nih.gov/articles/PMC10760010/
• https://academic.oup.com/gigascience/article/doi/10.1093/gigascience/giad043/7199403
• https://github.com/leelabsg/overlapGWAS