MTAG in overlapping twin sample

[rogercompte]

Hello,

I am thinking of using MTAG to meta-analyse (not with equal heritability nor perfect genetic correlation) different traits from the same condition/disease. I am working with two cohort data with the same different traits. One of them is based on twins with around 1000 participants and the other is a general population cohort with around 2500 participants. The idea is to (1) run GWAS in each cohort for each trait, (2) then join (meta-analyse) the GWAS results for each trait independently on both cohorts using a "normal" meta-analysis configuration, and at the end (3) use MTAG to join the different traits. GWAS for the twins cohort is performed with mixed models accounting for relationship of the subjects (using kinship matrix).

The first question is: can I use MTAG in such condition where there is a strong relatedness which has been taken into account in the GWAS and complete sample overlap? (with max-FDR included)

The second is: can I use MTAG that relies on LD regression in a small sample such this (<5000)?

The third is: a few of the traits I plan to use have also some GWAS run in other cohorts. If I include them on the first "normal" meta-analysis (2), the sample size will be actually very different along traits and thus each trait will be somehow biased to the traits with larger samples. I know you mentioned that on the paper, but just to clarify, is this bias too big so better keep just the two cohort data with full traits?

Hope the question is clear and thank you in advance!

Best, Roger

[paturley]

Hi Roger,

I'm a little confused about what you are proposing, but here are a few responses.

MTAG starts running into trouble when the power of the input GWAS are low because the parameters it estimates from LDSC can be way off in small samples. I generally tell people that if they want to use fully-flexible MTAG, it's best if the mean chi2 statistic in the GWAS is at least 1.1, though I've seen people use it for mean chi2 as low as 1.05.

The especially finicky parameter seems the genetic correlation parameter, so if you can make assumptions about the rg between the phenotypes, you'll have estimates that are much more stable.

In general, having GWAS samples that are largely overlapping is fine, even if they are based on samples of highly related individuals. I didn't totally follow your proposal, but the types of things you are talking about there are things that MTAG is usually quite robust to.

Let me know if you have any other questions.

Best, Patrick

On Tue, Dec 7, 2021 at 11:46 AM rogercompte @.***> wrote:

Hello,

I am thinking of using MTAG to meta-analyse (not with equal heritability nor perfect genetic correlation) different traits from the same condition/disease. I am working with two cohort data with the same different traits. One of them is based on twins with around 1000 participants and the other is a general population cohort with around 2500 participants. The idea is to (1) run GWAS in each cohort for each trait, (2) then join (meta-analyse) the GWAS results for each trait independently on both cohorts using a "normal" meta-analysis configuration, and at the end (3) use MTAG to join the different traits. GWAS for the twins cohort is performed with mixed models accounting for relationship of the subjects (using kinship matrix).

The first question is: can I use MTAG in such condition where there is a strong relatedness which has been taken into account in the GWAS and complete sample overlap? (with max-FDR included)

The second is: can I use MTAG that relies on LD regression in a small sample such this (<5000)?

The third is: a few of the traits I plan to use have also some GWAS run in other cohorts. If I include them on the first "normal" meta-analysis (2), the sample size will be actually very different along traits and thus each trait will be somehow biased to the traits with larger samples. I know you mentioned that on the paper, but just to clarify, is this bias too big so better keep just the two cohort data with full traits?

Hope the question is clear and thank you in advance!

Best, Roger

— 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/151, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5ONSKVZ52BJ5CDGRBLUPY25TANCNFSM5JRW4TZQ . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

[rogercompte]

Hi Patrick,

Sorry if I was not clear. Just for sake of clarification, the proposal is first to run single-trait GWAS of different traits in different cohorts and next perform inverse-variance-weighted meta-analysis, trait-wise, using the GWAS summaries of each trait in the different cohorts (e.g meta-analysis using GWAS summaries for trait 1 in all cohorts). This step is needed since raw data from the different cohorts cannot be merged. Finally use MTAG to join the results of each trait inverse-variance-weighted meta-analysis altogether.

The main question was if MTAG will perform well with GWAS statistic summaries coming from GWAS done on highly related subjects, like twins, where relatedness has been already accounted on the GWAS analysis (using kinship matrix); which you already answered.

In terms of the genetic correlation, the plan is to use different traits from a complex disease. Correlation of these traits is not perfect, for which I mean, having the disease does not mean having all traits but probably some of them. So even we expect some genetic correlation (that is the reason why we are combining the GWAS of the different traits) we don't think it is perfect (no --perfect_gencov flag).

I hope now is clearer and don't present any inconsistencies with the use of MTAG. Many thanks again.

Best wishes, Roger