I am reporting some cross-phenotype genetic correlations (LDSC method) that are statistically significant after multiple test correction. Peer reviewers have asked us to report top SNPs that might contribute to these correlations. I understand that genetic correlations reflect a genome-wide effect, but I want to satisfy reviewers/readers. Has the field arrived at a best practice for identifying/prioritizing SNPs/regions that mediate a significant genetic correlation? I have a few ideas below, but hope to get feedback from the community of experts.
1) It would be simple to threshold the 2 sets of GWAS results at some p-value (arbitrary threshold?) and compare the overlap of the two results. The surviving results could be clumped to LD-independent markers and nearby genomic features could be reported. Simple, but arbitrary selection.
2) Meta-analysis has been a common approach for combining data across phenotypes. We might see some signals persist or become more significant if a SNP is associated with both phenotypes. We also might see some associations driven by only one phenotype, so this method isn't exactly specific to SNPs contributing to correlation. Overall, this approach doesn't seem to answer the exact question I'm asking. From my cursory understanding, an MTAG analysis also wouldn’t exactly answer my question.
3) pHESS is a new method that seems promising for addressing my question. https://www.biorxiv.org/content/biorxiv/early/2016/12/08/092668.full.pdf; however, this approach does not control for ancestry stratification as nicely as the LDSC method.
As an extension on this question: Have we arrived at methods for assessing whether the SNPs/regions mediating a genetic correlation are significantly enriched for biological pathways/functions? The approach laid out here seems relevant, but I haven't had a chance to read the article (https://www.biorxiv.org/content/early/2017/03/07/114561)
Usually people analyse the variants for effect on genes, using something like the VEP (http://www.ensembl.org/info/docs/tools/vep/index.html). You would prioritise loss-of-function effects on genes over other consequences. You would analyse the genes for known phenotypes and GO terms to see if the genes' functions seem like they could reasonably be associated with the phenotype you're looking at. You can also analyse the allele frequencies in the 1000 Genomes and gnomAD populations, to see if the frequencies there are sensible compared to your phenotype: if it's a rare phenotype then the variants should also be rare, if it's common they should also be common.
Hi Daniel,
you mentioned only the tools to list important SNP/variants. I think you should add some pathways analysis as IPA or its free counterpart DAVID to select SNP and of course neighboring genes to go further the analysis and get closer to the pathology.
fred
Indeed, interesting question. Here are my 2 cents. Alternatively you could do the following. Choose one trait (presumably the trait of interest, i.e. "your" trait), and correlate that with the other traits, but: remove the top variants from 'your' data per p-value bin. For instance, exclude all variants with p
So far I've tried the pHESS software, but I'm finding it only identifies robust local correlations when phenotypes show a modest (> 0.2) genome-wide correlation and both are relatively well-powered (>50k subjects). The estimates and error bars for genome-wide correlation vary considerably from LDSC.
Additionally, I was pointed toward these resources by moderators on the LDSC user group forum:
- Bolormaa et al., 2014, PLoS Genet
- CPASSOC from Zhu et al., 2015, AJHG (see also Park et al, 2016, PLoS One)
- Binary Effects model from Han & Eskin, 2012, PLoS Genetics
- RE2C from Lee et al., 2017, Bioinformatics (slightly different angle on parsing shared vs. heterogeneous effects across cohorts)
- Badges
- Science topic
- Similar topics
- Bioinformatics
- Bioinformatics and Computational Biology