Meta-analytical methods for latent factor GWAS

[Johan Zvrskovec]

Hi,

Did you investigate any meta-analytical methods before settling on associating SNPs with latent factors through SEM? Is there anything you know of that would discourage their use as an alternative to what is achieved through userGWAS() in genomic SEM?

The reason I am asking is that I have been looking at doing exactly this when the genomic SEM userGWAS method got very slow for a larger number of GWAS and variants (eg 25 GWAS and the ~9M 1KG variants).

[agro...@gmail.com]

Hi Johan, 

There will be different opinions on this, but I would describe two main advantages of running a GWAS through Genomic SEM:

1. For 25 traits, it may be that possible representations of the data are more nuanced than a single factor  and/or that certain traits cluster together more than others. Genomic SEM allows you to formally model and quantify this, while most other meta-analytic GWAS methods are going to equally weight the traits being meta-analyzed and to my mind are often more suitable for meta-analyzing across different cohorts for the same trait. Of course this is a blanket statement, and other methods are (MA-GWAMA, N-GWAMA, MTAG) are designed to include different traits but could be said to have more of a goal of boosting power for a particular trait, as opposed to the goal of Genomic SEM to investigate shared and unique sources of genetic variation across a set of traits. 

2. One of the major benefits of multivariate GWAS in Genomic SEM is the estimation of Q_SNP, which allows you to guard against false positives that are identified as acting through the factor but are in factor specific to a particular trait or set of traits. As a particular case example, if you meta-analyzed a bunch of substance use traits you might find that variants in the ADH1B gene are significant in your meta-analysis, while Genomic SEM would likely identify these variants as significant for Q_SNP since they are unique to alcohol pathways. 

Hope this helps!

Best,

  Andrew

[Källberg Zvrskovec, Johan]

Hi Andrew,

 

Thank you for the clarifications. What I should have said was a weighted meta-analysis, weighted by the factor structure obtained from the genetic covariance matrix and how this compares to the Genomic SEM multivariate GWAS. I hoped that you may have considered such an approach since you use similar concepts (Q_SNP) as in a random-effect meta-analysis and cited:

http://journal.frontiersin.org/article/10.3389/fpsyg.2014.01521/abstract

as a source for Genomic SEM. If you did try similar methods we were wondering why they were disregarded in favour of the SEM multivariate GWAS.

 

Would a weighted random-effect meta-analysis Q statistic serve the same purpose as the Genomic SEM Q_SNP?

 

 

Best,

Johan

 

From: genomic-...@googlegroups.com <genomic-...@googlegroups.com> on behalf of agro...@gmail.com <agro...@gmail.com>
Date: Tuesday, 2 November 2021 at 17:43
To: Genomic SEM Users <genomic-...@googlegroups.com>
Subject: Re: Meta-analytical methods for latent factor GWAS

You don't often get email from agro...@gmail.com. Learn why this is important

--
You received this message because you are subscribed to a topic in the Google Groups "Genomic SEM Users" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/genomic-sem-users/VNnmj_e8zGo/unsubscribe.
To unsubscribe from this group and all its topics, send an email to genomic-sem-us...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/genomic-sem-users/5ed84733-9511-4e29-aec4-f1d22bf40022n%40googlegroups.com.

[Elliot Tucker-Drob]

Hi Johan,

You are asking about something that I would only recommend for a very advanced user. A weighted meta-analysis still assumes the same population average effect size across all studies, and simply weights by the precision of the estimates (the inverse standard error squared). To further weight by the factor loadings (which is achieved by multiplying weights together; i.e. the loadings times the inverse sampling variance) would simply shift the mean effect size closer to that for the phenotypes with higher loading estimates but it would not give you the effect on the factor, and its heterogeneity statistic (tau squared, the significance of which is indexed with the meta-analytic test statistic Q) would not reflect the extent to which the SNP deviates from the factor expectations, but simply reflect whether the SNP effects are heterogenous across traits (which we would expect them to be if the factor loadings are uneven across traits). To approximate a common factor GWAS, the closest you could get would be to rescale the SNP effects and SEs by the reciprocal factor loadings (from the no-SNP model) and then conduct an inverse variance weighted random effects meta-analysis using the rescaled SNP estimates. Then, the heterogeneity would indeed reflect deviations from the factor model expectations, and the effect size estimate would represent the SNP effect on the factor. Again, you'd really need to know what you're doing, and I'd want to make sure that this was benchmarked against a commonfactorGWAS analysis and against simulated data. It would also only work for a simple common factor GWAS: it would not be straightforward to adapt to other types of models. To add, I would strongly advise against attempting to conduct any meta-analysis of this sort using Z statistics. Effect sizes are key here, so the inverse variance weighted approach is best suited.

Please note that sample overlap or shared stratification across GWAS traits is automatically accounted for when conducting multivariate GWAS in Genomic SEM. You would need to further adapt your meta-analysis to accommodate such potential confounds if conducting outside of Genomic SEM (see, e.g. https://www.nature.com/articles/s41588-018-0320-8)

I want to emphasize that we have prioritized maintining Genomic SEM as a flexible modeling framework, rather than a tool to apply a. specific model. In order to preserve the flexibility of Genomic SEM, we have not optimized the code to reduce runtime for a specific model. We do realize that commonfactorGWAS is often the model of interest for people, so it is possible that in the future we will create a lightning quick version of that function. That said, there is already a tendency for some people to think of SEM as single common factor modeling, and we do not want to reinforce that impression. It is much broader and is well-suited to many different applications that have nothing to do with common factors.

Elliot

You received this message because you are subscribed to the Google Groups "Genomic SEM Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to genomic-sem-us...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/genomic-sem-users/AM7PR03MB6499FFEF76CB39F577A2B194F08C9%40AM7PR03MB6499.eurprd03.prod.outlook.com.

[Källberg Zvrskovec, Johan]

Hi Elliot,

 

Thank you. I first hoped we could apply standard meta-analytical methods without simulations and benchmarks, but judging from your answer and also Gerome’s input, is that it is not straightforward and will turn my project into a lengthy methods project.

 

I was actually interested in applying the meta-analysis on more complex factors/models than a common factor, which was another reason why I wanted to investigate the meta-analytical approach from the start. The Genomic SEM multivariate GWAS does not work as well for uncorrelated factor models.

It would be super helpful to have your, Andrew’s and Michel’s input on the actual formulas. When ready, I can send some material over in a private message if you are interested and have the time for it.

 

Best,

Johan