Scales of polygenic risk score distributions

[johannfaouzi]

Hi Florian,

I have a question regarding the scales of the polygenic risk score distributions computed for a given population.

In short, I have computed PRS for several phenotypes and I'm interested in their potential association with measured phenotypes. To do so, I use linear mixed effects models (I have longitudinal data) with one model for each PRS, and I check the sign and the p-value of the coefficient for the PRS.

Now I'm wondering if the actual value of the coefficient is interpretable, because the p-value only indicates the significance but not the effect size. I have computed all the PRS using the same approach (LDpred2-auto) and I have noticed that the scales for different PRS may be very different.

So I'm wondering if the scales of PRS distributions (for a given population) are comparable between different PRS or not for whatever reasons (summary statistics, chains used to compute the final PRS, etc.).

Best, Johann

[privefl]

This is not something I have checked, so I don't know, sorry.

[johannfaouzi]

No problem, thank you for your reply!

[johannfaouzi]

I just checked the estimated standard deviations of the phenotypes with

sd_phenotype <- min(with(info_snp, beta_se * sqrt(n_eff) * sqrt(0.5)))

and the values are quite different. I'm going to check if, after dividing the PRS by the estimated standard deviation of each phenotype, the scales are more similar.

[privefl]

My guess would be that the SD of the PGS is also proportional to the R^2.

[johannfaouzi]

On a log-log scale the plot is quite linear, but the coefficient (thus the power) is a bit far from 1 (around 0.9).

In the linear scale there are quite a few outliers, which can be seen when I compute the 5 quartiles of the ratio distribution (0.001040, 0.114378, 0.158402, 0.244434, 0.936943).

I may give up on this for the moment.

[privefl]

Can you color by R^2 maybe?

[johannfaouzi]

By R^2 you mean the SNP heritability reported in the GWAS (or estimated by the ldsc function)?

[privefl]

Not h^2, but R^2, the variance explained by the PGS. But you can also compare with h^2.

[johannfaouzi]

I'm computing mostly PGS for phenotypes that are not measured (it's an exploratory analysis), so I think that I can't compute the variance explained by the PGS.

[johannfaouzi]

Here is the scatter plot with the h^2 values used as colors.