This is a test version of Biostars. For the public version, visit https://www.biostars.org.
How do you set the GWAS threshold?

I am currently conducting a GWAS using SNPs and genomic data obtained from RNA-Seq data. I have managed to generate a Manhattan plot, but I am struggling with setting the threshold. How do you all set or calculate the threshold? Please share your ideas. The dataset consists of 68 samples and 22,996 SNPs.

gwas rna-seq threshold snp
• 205 views • link
written by CaERu _no_ korokoro • CaERu _no_ korokoro

Just as a word of warning, I'd be quite skeptical about genotypes coming from RNA-seq data. At a very minimum you need to discard any SNPs that could actaully be RNA editing sites. Also be aware that a SNP in RNA-seq could be due to a combination of missmapping of reads from gene A to gene B, and differential expression of gene B.

• 0 views • link
written by i.sudbery i.sudbery

1 answer

The classic threshold for a GWAS study is p < 5x10^-8, but this is for a design including millions of SNPs, its likely that a more lenient threshold would be acceptable with only 23,000 SNPs.

This threshold was set by following three seperate lines of arguement, that all happened to point to the same threshold.

  1. At the most simple level 5x10^-8 is the bonferroni correction of a threshold of 0.05 corrected for testing 1,000,000 SNPs. So you could just take 0.05/22996 ~= 2.2x10^-6.

  2. However, in GWAS, SNPs are not independent. They come in haplotypes. So in a human GWAS, the number of independent tests is not really just the number of SNPs. Studies that tried to calcualte the "effective" number of independent test in the human genome also landed on a number that was of the order of p<10^-8 to p<10^-9 (1)

  3. The final line was a baysian line of reasoning. If you assume that there are 10 to 100 genuinely associated SNPs amoungst the billions of bases in the human genome, the prior probability of any one of them being associated with a trait is very small. This means that the Bayes Factor for association needs to be very high to overcome these small prior probabilities (2) . If you convert those bayes factors back to p-values (for the human genome) using reasonable assumptions, you get p-value thresholds in the same order of magnitude as other methods. You could investigate the proccedures in (2) to get your own threshold for your situation.

• 0 views • link
written by i.sudbery i.sudbery

Thank you very much for your detailed and insightful explanation. It was incredibly helpful to understand the theoretical background of the thresholds. My material is a plant F2 population, I suspect that even the Bonferroni-corrected threshold (p=2.2×10 6p=2.2×10 6) might still be too conservative for my data. I understand that QTL mapping is generally more appropriate for an F2 population. However, QTL analysis has already been conducted in our previous studies. I am now applying GWAS to gain more detailed insights and potentially higher resolution. Ultimately, I do not intend to rely solely on GWAS to identify the causal genes. Since I have RNA Seq data, I am planning to take a comprehensive approach by integrating the GWAS results with multiple perspectives, such as Differentially Expressed Gene (DEG) analysis and Weighted Gene Co expression Network Analysis (WGCNA).

• 0 views • link
written by CaERu _no_ korokoro • CaERu _no_ korokoro

Log in to answer this question.

More posts like this