Questions about the effect size inflation of MTAG

[test12138jooh]

Dear professor， We conducted an analysis using MTAG. Overall, the correlation between the effect sizes from MTAG and the original GWAS is quite good, r=0.67; however, we found that the effect sizes of some loci have changed significantly. The previous explanation was that the effect sizes estimated by MTAG were within the confidence intervals of the original GWAS estimates. https://github.com/JonJala/mtag/issues/209

However, we have found that many loci do not meet this criterion. Here are some examples.

RAW phenotype1 GWAS summary：beta=-0.05；SE=0.01616；P=0.001324 MTAG phenotype1 summary：beta=-0.12；SE=0011；P=2.04E-25

Thanks again.

Best, JOOH

[paturley]

What are the betas and standard errors for the other phenotypes included in MTAG and what is the estimated Omega that the MTAG log file reports?

On Wed, Aug 7, 2024 at 10:09 AM test12138jooh @.***> wrote:

Dear professor， We conducted an analysis using MTAG. Overall, the correlation between the effect sizes from MTAG and the original GWAS is quite good, r=0.67; however, we found that the effect sizes of some loci have changed significantly. The previous explanation was that the effect sizes estimated by MTAG were within the confidence intervals of the original GWAS estimates. https://github.com/JonJala/mtag/issues/209 http://url

However, we have found that many loci do not meet this criterion. Here are some examples.

RAW phenotype1 GWAS summary：beta=-0.05；SE=0.01616；P=0.001324 MTAG phenotype1 summary：beta=-0.12；SE=0011；P=2.04E-25

— Reply to this email directly, view it on GitHub https://github.com/JonJala/mtag/issues/218, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5JZHAOB3ZWRLTNDAKLZQIS75AVCNFSM6AAAAABMEQBIGOVHI2DSMVQWIX3LMV43ASLTON2WKOZSGQ2TGNJYGY2TMNA . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[test12138jooh]

RAW phenotype2 GWAS summary：beta=-0.0416；SE=0.0031；P=2.296e-41

And all phenotypes have been normlized before the gwas.

Attached is the log file . Thanks!

tmp.log

[paturley]

So in this case, it looks like you have a really high powered trait 2 with a low powered trait 1 and a moderate correlation between them. So MTAG is putting a lot of weight on the trait 2 effect size. If the MTAG assumptions are satisfied, this should be fine and your results should be valid, but as we describe in the MTAG paper, in this particular case MTAG results will be shaded towards the effect sizes of the higher powered trait. You should be able to see this if you estimate the genetic correlation between the GWAS results for trait 1 and the GWAS results for trait 2 vs the MTAG results for trait 1 versus the GWAS results for trait 2. I suspect that the rg in the latter case will be much higher. (A bit higher is expected just due to the way that MTAG works, but substantially higher may be indicative of a problem.)

On Wed, Aug 7, 2024 at 10:24 AM test12138jooh @.***> wrote:

RAW phenotype2 GWAS summary：beta=-0.0416；SE=0.0031；P=2.296e-41

Attached is the log file . Thanks!

tmp.log https://github.com/user-attachments/files/16529964/tmp.log

— Reply to this email directly, view it on GitHub https://github.com/JonJala/mtag/issues/218#issuecomment-2273603203, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5IY4OYIFSJ23ZXIPLTZQIUYFAVCNFSM6AAAAABMEQBIGOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDENZTGYYDGMRQGM . You are receiving this because you commented.Message ID: @.***>

[test12138jooh]

Thanks for your reply.

However, I am still confused. Even if, as you mentioned, MTAG will lean towards pheno2, the example data I showed indicates that the beta values for the two phenotypes are close. Yet, after applying MTAG, the beta values inflated twofold, which is higher than the effect sizes for both phenotypes in the original GWAS.

[paturley]

True. Do you know what the variance was for phenotype 1 before you ran the GWAS? MTAG units are always in the units of the standardized phenotype.

On Wed, Aug 7, 2024 at 10:47 AM test12138jooh @.***> wrote:

Thanks for your reply.

However, I am still confused. Even if, as you mentioned, MTAG will lean towards pheno2, the example data I showed indicates that the beta values for the two phenotypes are close. Yet, after applying MTAG, the beta values inflated twofold, which is higher than the effect sizes for both phenotypes in the original GWAS.

— Reply to this email directly, view it on GitHub https://github.com/JonJala/mtag/issues/218#issuecomment-2273655312, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5JI2I5S7327PSNM4DTZQIXPRAVCNFSM6AAAAABMEQBIGOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDENZTGY2TKMZRGI . You are receiving this because you commented.Message ID: @.***>

[test12138jooh]

yeah, I have performed the z-score transformation before I ran the GWAS, so the SE phenotype is 1 and mean is 0.

[test12138jooh]

So in this case, it looks like you have a really high powered trait 2 with a low powered trait 1 and a moderate correlation between them. So MTAG is putting a lot of weight on the trait 2 effect size. If the MTAG assumptions are satisfied, this should be fine and your results should be valid, but as we describe in the MTAG paper, in this particular case MTAG results will be shaded towards the effect sizes of the higher powered trait. You should be able to see this if you estimate the genetic correlation between the GWAS results for trait 1 and the GWAS results for trait 2 vs the MTAG results for trait 1 versus the GWAS results for trait 2. I suspect that the rg in the latter case will be much higher. (A bit higher is expected just due to the way that MTAG works, but substantially higher may be indicative of a problem.) … On Wed, Aug 7, 2024 at 10:24 AM test12138jooh @.> wrote: RAW phenotype2 GWAS summary：beta=-0.0416；SE=0.0031；P=2.296e-41 Attached is the log file . Thanks! tmp.log https://github.com/user-attachments/files/16529964/tmp.log — Reply to this email directly, view it on GitHub <#218 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5IY4OYIFSJ23ZXIPLTZQIUYFAVCNFSM6AAAAABMEQBIGOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDENZTGYYDGMRQGM . You are receiving this because you commented.Message ID: @.>

Besides, the raw gwas correlation was 0.37, and the genetic correlation of mtag result was 0.45，it seems normal? I calculated with LDSC.

[test12138jooh]

I found a parameter "--std_beta", Should I use this in my command? I noticed you have explained it before： SD of the phenotype per allele change (no --std_betas) SD of the phenotype per standard deviation of the genotype (use --std_betas)

But I still fill confused about that; I think the genotype or the allele sd seems the same meanning in GWAS？Besides，I found that if I use this parameter, the beta value I got seems to get closer to the raw beta( By the way, the raw beta was from results of plink using linear association)?

2. I also feeel confused about the purpose of using weights to adjust the beta? It seems the weights are calculated by the MAF of the variant, is it right?

Do you have any suggestions about that? Thanks!

[paturley]

Hi,

Sorry I've disappeared for a bit. I'm traveling this week, so I'm going to have pretty limited availability to troubleshoot.

I think you should not use the std-betas option. Virtually all GWAS are reported in allele count units. I'm not even sure why we added that option.

Have you tried running MTAG with the same options as you do with 2 phenotypes, but just running it one phenotype at a time? If the units of the GWAS results and the MTAG results is the same, the betas should be essentially the same (though the SEs will be different).

Patrick

On Sun, Aug 11, 2024 at 9:37 AM test12138jooh @.***> wrote:

I found a parameter "--std_beta", Should I use this in my command? I noticed you have expalined it before： SD of the phenotype per allele change (no --std_betas) SD of the phenotype per standard deviation of the genotype (use --std_betas)

But I still fill confused about that; I think the genotype or the allele sd seems the same meanning in GWAS？ Besides，I found that if I use this parameter, the beta value I got seems to get closer to the raw beta(My raw beta was from results of plink using linear association)? Do you have any suggestions about that? Thanks!

— Reply to this email directly, view it on GitHub https://github.com/JonJala/mtag/issues/218#issuecomment-2282764143, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5MNJELYWA7L2M2KAJ3ZQ5SKJAVCNFSM6AAAAABMEQBIGOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDEOBSG43DIMJUGM . You are receiving this because you commented.Message ID: @.***>