JonJala / mtag

Python command line tool for Multi-Trait Analysis of GWAS (MTAG)
GNU General Public License v3.0
172 stars 55 forks source link

Meta-analyze MTAG outputs? #133

Open humanpaingeneticslab opened 3 years ago

humanpaingeneticslab commented 3 years ago

Hi Everyone,

this is a question rather than an issue;

Can one conduct a meta-analysis of MTAG outputs? Is this viable, or it would be like double-counting (as the SNPs put forward in the meta were already boosted by MTAG)?

Thanks!

paturley commented 3 years ago

Hello.

Just to make sure I understand, you are asking if it is OK to combine a set of summary statistics using MTAG and using the resulting sets of MTAG summary statistics to combine them again using meta-analysis? Generally, this will not be OK. When you combine sets of summary statistics with MTAG, the error between the sets of MTAG summary statistics is correlated, so a standard meta-analysis that doesn't take this correlation into account would have standard errors that are too tight.

If instead what you meant was that you wanted to take a set of MTAG summary statistics and meta-analyze them with a different set of GWAS summary statistics that didn't have any sample overlap with any of the cohorts used in the MTAG analysis, then I think that should be OK.

Does that answer your question?

Best, Patrick

On Mon, Jun 14, 2021 at 6:43 AM unepanthere @.***> wrote:

Hi Everyone,

this is a question rather than an issue;

Can one conduct a meta-analysis of MTAG outputs? Is this viable, or it would be like double-counting (as the SNPs put forward in the meta were already boosted by MTAG)?

Thanks!

— 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/133, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5LORMAJQ2BEZKX7ZQ3TSXMLVANCNFSM46U7EYWQ .

humanpaingeneticslab commented 3 years ago

Hi Patrick,

Thank you for your prompt response.

I think I'd fall in your first setup, although let me try to be more explicit. Suppose I have three summary GWAS from three independent cohorts of related but not identical phenotypes; let's call them S1, S2 and S3. I can then use MTAG to analyse them simultaneously, and MTAG will then provide adjusted summary GWAS for each; let's call them A1, A2 and A3. Right now I'm doing a meta-analysis of S1, S2 and S3. I was wondering if instead I meta-analyzed A1, A2 and A3 I would have better genetic signal for the SNPs pleiotropic for my phenotypes. In the paper, the follow-up analyses are on a per-phenotype basis, but not from a "combined" analysis of MTAG-adjusted results.

Very Best, Marc.

paturley commented 3 years ago

Hi Marc,

If I understand correctly what you are doing, in your case, the error of A1, A2, and A3 will be correlated. This is because there is sample overlap between the A1, A2, and A3 samples (since they are all based on the S1, S2, and S3 samples). So meta-analyzing A1, A2, and A3 will result in standard errors that are too tight. If you want to produce summary statistics that give the combined average effect of each SNP across your three phenotypes (so that pleiotropic SNPs with concordantly-signed effects are more likely to be significant), then you are better off combining S1, S2, and S3 since there is no sample overlap between them.

Am I understanding your question correctly?

Best, Patrick

On Mon, Jun 14, 2021 at 10:38 AM unepanthere @.***> wrote:

Hi Patrick,

Thank you for your prompt response.

I think I'd fall in your first setup, although let me try to be more explicit. Suppose I have three summary GWAS from three independent cohorts of related but not identical phenotypes; let's call them S1, S2 and S3. I can then use MTAG to analyse them simultaneously, and MTAG will then provide adjusted summary GWAS for each; let's call them A1, A2 and A3. Right now I'm doing a meta-analysis of S1, S2 and S3. I was wondering if instead I meta-analyzed A1, A2 and A3 I would have better genetic signal for the SNPs pleiotropic for my phenotypes. In the paper, the follow-up analyses are on a per-phenotype basis, but not from a "combined" analysis of MTAG-adjusted results.

Very Best, Marc.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/JonJala/mtag/issues/133#issuecomment-860737896, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5KPA5WW5D4VL5H7Q4DTSYH7NANCNFSM46U7EYWQ .

humanpaingeneticslab commented 3 years ago

Hi Patrick,

OK, yes. I think my main problem is that I seem to struggle to understand the difference between a meta-analysis and MTAG. If I perform a meta-analysis, the hope is that SNPs with concordantly-signed effects will be more likely to be significant. When I use MTAG, the SNPs with concordantly-signed effects are "boosted" in each summary GWAS, and that's the output of MTAG?

Thanks again!

paturley commented 3 years ago

You can think of MTAG as a generalization of meta-analysis. In fact, if the samples don't overlap, the traits are perfected genetically correlated, and they have the same heritability, MTAG is equivalent to IVW meta-analysis.

On Mon, Jun 14, 2021 at 11:05 AM unepanthere @.***> wrote:

Hi Patrick,

OK, yes. I think my main problem is that I seem to struggle to understand the difference between a meta-analysis and MTAG. If I perform a meta-analysis, the hope is that SNPs with concordantly-signed effects will be more likely to be significant. When I use MTAG, the SNPs with concordantly-signed effects are "boosted" in each summary GWAS, and that's the output of MTAG?

Thanks again!

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/JonJala/mtag/issues/133#issuecomment-860759195, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5N2VOKW5DPT6V65FCLTSYLEJANCNFSM46U7EYWQ .

anglixue commented 1 year ago

Hi @paturley

I am reading this old thread. I was wondering for a typical MTAG result, are the SNP effects still marginal? or they became joint effects which takes LD into account?

If like you said, under the scenario where the samples don't overlap and perfected correlated/same heritability and MTAG is equivalent to IVW meta-analysis, I believe the MATG effects are still marginal?

paturley commented 1 year ago

Yes, that's right. The output of MTAG are marginal effects with the same interpretation as what you'd get from a GWAS.

On Wed, Sep 20, 2023 at 3:51 AM Angli Xue @.***> wrote:

Hi @paturley https://github.com/paturley

I am reading this old thread. I was wondering for a typical MTAG result, are the SNP effects still marginal? or they became joint effects which takes LD into account?

If like you said, under the scenario where the samples don't overlap and perfected correlated/same heritability and MTAG is equivalent to IVW meta-analysis, I believe the MATG effects are still marginal?

— Reply to this email directly, view it on GitHub https://github.com/JonJala/mtag/issues/133#issuecomment-1727162875, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5NCBRAHFEJAK2VIWYLX3KN7HANCNFSM46U7EYWQ . You are receiving this because you were mentioned.Message ID: @.***>