I recently got involved into a project that involves calculation of genetic risk scores (GRS) for a set of important SNPs. My question is about weighted GRS and it is as follows,
PredictABEL calculates weighted GRS as sum of allele counts weighted with beta coefficients. Does it matter whether or not the beta coefficients are standardized (resulted from standardized variables)?
Some suggest rescaling weighted GRS by multiplying GRS with number_of_SNPs/sum_of_beta_coefficients to make it comparable to unweighted GRS. Could someone help me understand how and when that is beneficial?
In unweighted, the model makes the assumption that all markers are of equal importance, which, when applied to genetic susceptibility, doesn't make much sense unless you have already carefully honed down on a select few markers in a test panel that do indeed carry the same importance / association to disease / amount of information.
In weighted, the amount of information that each marker contributes, i.e., it's importance, is gauged via its beta coefficient from regression. Look up the bayesmodel() R function, which allows you to do something similar. In English terms, this means that a marker that has a weak coefficient will have to provide a major amount of evidence in the test data in order to prove its worth.
I would always favour weighted unless you had a very good reason not to choose it.
Thanks for your reply! But that still does not answer my question. I am sorry if I was not clear enough.
My question is how scaling of weighted GRS matters. Weighted GRS calculated by predictABEL is not scaled. However, some authors, however, suggest rescaling weighted GRS by dividing it with mean of beta coefficients.
I tried logistic regression with both scaled and not scaled weighted GRS and results are different. So, I want to know what is correct way?
Thanks for posting the study! - looks interesting. From what I can see, scaling the weighted GRS is mainly to allow for a better comparison to the unweighted GRS. I can only imagine that weighting the GRS alters the distribution of the scores substantially, such that a comparison to the unweighted scores is not possible unless you rescale them to a distribution that matches that of the unweighted.
In conclusion, just looking at this study alone, it appears that re-scaling is only performed for the sake of comparison to unweighted scores, and is therefore not an essential step.
Thanks for your reply! But that still does not answer my question. I am sorry if I was not clear enough.
My question is how scaling of weighted GRS matters. Weighted GRS calculated by predictABEL is not scaled. However, some authors, however, suggest rescaling weighted GRS by dividing it with mean of beta coefficients.
I tried logistic regression with both scaled and not scaled weighted GRS and results are different. So, I want to know what is correct way?
Krishna
Hello!
Can you point me to the published work where it states that scaled weighted is better?
Kevin
Sure, in the following article, Lin, X., Song, K., Lim, N. et al. Diabetologia (2009) 52: 600. https://doi.org/10.1007/s00125-008-1254-y
They have used rescaled weighted GRS, but it is not clear to me how the rescaling helps.
I appreciate for your time and help!
Thanks for posting the study! - looks interesting. From what I can see, scaling the weighted GRS is mainly to allow for a better comparison to the unweighted GRS. I can only imagine that weighting the GRS alters the distribution of the scores substantially, such that a comparison to the unweighted scores is not possible unless you rescale them to a distribution that matches that of the unweighted.
In conclusion, just looking at this study alone, it appears that re-scaling is only performed for the sake of comparison to unweighted scores, and is therefore not an essential step.