Question: FarmCPU - how to explain the reported 'effect'?

0

Philipp Bayer ♦

**6.0k**wrote:When you run a GWAS using FarmCPU via GAPIT or MVP, you get an 'effect' reported per SNP (a number somewhere around -1 to +1, with a bunch of outliers).

I've looked at the papers and the GAPIT forum but I'm not understanding what the effect itself reports. Is it the effect of the minor allele? The major allele? The entire SNP in the model? And is that effect measured in the same unit as the associated phenotype, i.e., if I have a plant height phenotype around 50 cm, does a SNP with an effect of +5 lead to a 5cm higher plant in the model?

Hey Philipp, just on the sign of the effect (+ / -), it seems that you can fix the major allele to have 0 by specifying

`Major.allele.zero = TRUE`

. Thus, the given effect is always relating to the minor allele.Regarding the interpretation of the number, I'd have thought that it was merely the estimate / coefficient from the fitted model? Thus, the exponent of the effect should be the odds ratio? May need to confirm with the authors.

41kThank you Kevin! Yes, the coefficient makes much more sense - I'm using MVP currently which doesn't let me set

`Major.allele.zero`

, I will have to ask the authors!6.0kHello Phillip,

A better way to think about this would be if aa=0 aA=1 and AA = 2 which is typical numerical encodings used in GWAS. Then for the simplest case where you have 1 snp your model would be Pheno = b0 + b1*snp where b0 is the mean of the trait and b1 is the "effect"

So for your example with plant Height (PH) if the effect of the allele is 5 and the mean of the trait is 50 I can show how the model would look.

aa: PH = 50 + 5x0 = 50 aA: PH = 50 + 5x1 = 55 AA: PH = 50 + 5x2 = 60

I think the effect is the estimated effect of a gene in the Final fixed effect model of farmCPU. You have to be careful with effects, if you have very small MAF then the effect was calculated with a small fraction of individuals as 0. Also I would not read into the effect too much in my opinion. farmCPU is a GWAS model, not genomic selection model so the point of it is not necessarily to tell you how much of an effect a particular gene has but rather give you a good idea of what genes control a trait. It would guess if you do farmCPU for two populations with the same snp chip that the effect of a particular gene will not be the same (likely you wont even get the same hits either).

If you are concerned alot with the qtn effect try fitting either a ridge regression (rrBlup) or lasso regression (glmnet) model to your data and then pull out the effect for the qtn you found in farmCPU.

0