Hi philippa

commonly such omics analysis give huge data, leading the necessity to filter amount of results. try at 0.05 and see the results (I mean size of results and candidate gens you could have, even with missing data), decreasing step by step the pValue till the list is acceptable for further analysis.

fred

Dear Philippa,

Could you be more clear which GWAS data do you use, your own or from GWAS catalog? In case there are no target phenotype data, you have no clues (also metadata can be used) to reduce the amount of SNPs in the final list applying just the P value criteria. Generally, I would avoid using just the P value for such tasks as its statistical level is reached basically by the number of observations of the association between the phenotype and the SNP. So, I apply additionally criteria as SNP gene attribution (grouping by gene), chromosome localization (grouping by region), GWAS catalog trait or even linkage between them. Besides, dont forget that many of SNPs lye in non coding regions, so you may think about regulatory aspects.

Anyway just statistical approach is not enough, biological meaning also can be an aid.

Good luck.

Thank you very much for your advice.

I am using data from GWAS catalogue as the base dataset, checked for strand alignment, I was going to use either clumping or a lasso regression to account for linkage disequilibrium.

In the target data set (TwinsUK), I plan to first perform all QC checks as if I was going to run a GWAS.

I like the suggestion of using gene attribution/chromosome localisation, but I am unsure how to apply this on a practical level, please could you suggest a paper that would go into greater detail about these methods?

Thank you again for taking the time to adivse me, I really appreciate it.

Philippa

Dear Philippa,

Even the GWAS catalog data can be somehow filtered. I used the data to identify very few highly significant SNPs with obesity related traits. So, I extracted all related datasets and then filtered the data starting from studied population (Caucasians), type of the traits described (I was interested in lipids traits), chromosom location and gene association of the SNPs (excluding sexual chromosoms), then grouped the data by the trait, ranging them in increasing order of the P value, thus obtaining Genes and their highly associated to lipids traits in the first ranges. That was enough for me to chose SNPs for 9 genes. Now I am performing TaqMan genotyping to validate GWAS evidences on our population.

Previously, I used similar approach to analyze microarray data, grouping highly correlated to studied process genes and checking corresponding P value (it is another usage of the P value, different the one you would like to use). Then I looked for the genes (in your case SNPs) for which the association was already demonstrated through dedicated studies. It helped me to check if my rational returns me at least these genes. So I could identify couple of them, while the rest of the genes showed no significant (in my study) level. 

Another my experience included the usage of the Gene Ontology to group the genes by molecular process, biological function or protein cellular localization. And here we had to check the biological meaning, e.g. a hormone treatment should stimulate some of the metabolic processes (aminoacid transformation, protein synthesis etc).

So, these are just examples of the logic you may try. Still the question will be - what is the hypothesis? what the P value you use for testing a hypothesis?

The question you have described seems to be a theoretic one, until there is no biological meaning, eg, are you looking for the association between a locus, which includes a set of SNPs from different genes, which association is estimated by LD using any kind of regression, with a TRAIT?

And, as I can imagine once you use TwinsUK dataset, you will need to assess the contribution of these loci to a trait, givind a score. So, you will get a kind of impact estimation of SNP/SNPs/loci to a phenotype. In this case, P value will give just statistical "sure" about the score and cannot be used itself in score calculation.

Sorry writing too much and confusing you.

Kind regards,

Hi Philippa,

If you want to rely solely on discovery sample GWAS summary statistics for choosing the groups of SNPs to include in the score, you can run the analysis several times across a set of decreasingly stringent p-values, while accounting for LD and filtering for other parameters, such as r^2. You might expect to see an improvement in prediction as your threshold becomes more lenient (i.e. you are including increasingly larger sets of SNPs that are associated with the discovery trait), until you start to incorporate noise (i.e. you've included SNPs that are unassociated with the discovery trait into the score). I believe this might be the method to which Frederic was referring, above. 

A couple of recent manuscripts that have described this method are: Webb et al., 2017 and Bigdeli et al., 2017.

Hope this helps.

Thank you for the advice! I really appreciate it.

Best Wishes,

Philippa