No BIM Filter?

[MatthewMaher]

Hi- Thanks for PRScs* - great tools.

Question: I understand PRScs produces a PRS model composed of SNPs that are in the three-way intersection of the base-GWAS, the LD-panel, and the target cohort. And including that last one is critical to ensuring that a target cohort can be scored with 100% coverage (i.e. no missing variants).

But then I see general PRS-websites/portals where people post big PRS-models derived from PRScs (or other similar tools) for no particular target cohort. I'm supposing this usually works out okay because hopefully most target cohorts have good availability for the ~1.1M (or whatever it is) SNPs in the LD-ref panels, which limits what goes into the finished model, and thus the gap/loss due to missing SNPs would be minimal. Am I understanding that correctly?

So anyway, if I wanted to run PRScs like that - i.e. for no particular target cohort - and just rely on the filtering of the LD-panel, is there a shorthand way to do that (i.e. --nobim? ). It seems like PRScs requires the --bim_prefix option. Of course, I could create a fake BIM out of the base GWAS summary stats, and get the desired result. But that seems like jumping through a pointless hoop, which makes me wonder if I'm missing a simpler way?

Or maybe I'm misunderstanding the criticality of the BIM and the quality of all those non-target-specific models out there.

Thanks for any info/insight

[getian107]

Hi - Yes you can build PRS for no particular target cohort. There is some risk that if a variant is assigned a large effect size but isn't available in the target dataset the predictive accuracy would decrease. But HapMap3 variants are usually well imputed and for the vast majority of polygenic traits/diseases, missing a small number of variants in the target datasets has minimal impact on prediction performance.

PRS-CS does require a bim input. You can create a fake bim from the SNP info file that comes with the software and includes all the variants on the reference panel.

[MatthewMaher]

sounds good - thanks for the info.