search

JonJala / mtag

Python command line tool for Multi-Trait Analysis of GWAS (MTAG)
GNU General Public License v3.0
169 stars 54 forks source link

Low genetic correlated traits #138

Open zongchangli opened 3 years ago

zongchangli commented 3 years ago

Hello, My current project aims at weighting GWAS snp effects for trait1 with GWAS snp effects from trait2. The LDSC rg genetic correlations is about 0.1. In my scenario, do you think it's ok for using this software for analysis?

paturley commented 3 years ago

Hello Zongchan,

In the paper, we give a rule-of-thumb recommendation of rg>0.7, but it will depend on your application and sample size. I recommend you run the maxFDR calculation to see if you are willing to accept a risk of a FDR as high as what is reported.

Patrick

On Tue, Jul 20, 2021 at 9:51 PM Zongchang Li @.***> wrote:

Hello, My current project aims at weighting GWAS snp effects for trait1 with GWAS snp effects from trait2. The LDSC rg genetic correlations is about 0.1. In my scenario, do you think it's ok for using this software for analysis?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/JonJala/mtag/issues/138, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5PVZGH3I5KR5S2BC5TTYYR3JANCNFSM5AW7KFPA .

zongchangli commented 3 years ago

Thanks

divbru commented 1 year ago

Hi Patrick, could you please elaborate on what you mean by "depend on application and sample size"? Additionally, do you have any suggestions/guidelines (apart from running maxFDR, as it seems to be extremely computationally intensive) to improve inference when considering phenotypes with moderate/low r_g (~0.4/0.2) ?

paturley commented 1 year ago

So MTAG is robust to low rg as long as its assumptions are met. So for example, if you truly believe that the effect sizes follow a bivariate normal distribution with correlation 0.4, then MTAG should be just fine. Also, MTAG tends to be fairly robust if the mean chi2s are comparable across phenotypes. (You can see this if you look at the simulation results of the original paper.) You get in big trouble when there are SNPs that affect trait1 but not trait2, the genetic correlation is high (but not perfect) between the traits, and the sample size is much larger for trait1. In that case, you may get some SNPs that MTAG says are statistically significant for trait2 when they are actually just contamination from trait1.

One thing to think about though is that if the rg is low, then MTAG will not share very much information across traits. This means that the risk of false positives is low (which I suppose is good) but it also means that MTAG is not going to boost power much (which is less helpful).

On Thu, Mar 9, 2023 at 12:56 PM Divya @.***> wrote:

Hi Patrick, could you please elaborate on what you mean by "depend on application and sample size"? Additionally, do you have any suggestions/guidelines (apart from running maxFDR, as it seems to be extremely computationally intensive) to improve inference when considering phenotypes with moderate/low r_g (~0.4/0.2) ?

— Reply to this email directly, view it on GitHub https://github.com/JonJala/mtag/issues/138#issuecomment-1462511162, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5LJG2I545GTA77ZACTW3IKWDANCNFSM5AW7KFPA . You are receiving this because you commented.Message ID: @.***>