Finemap

[jameskozubek]

Thanks for the code.
I am stuck at this step 03_run-finemap.sh
I guess I should download this?
http://www.christianbenner.com/

Correct. You will also need LDStore, as well as bgenix, which is part of bgen, found here https://enkre.net/cgi-bin/code/bgen/dir?ci=trunk.

Also, I just realized the Makefile is in the wrong directory (02_pair-genes-to-snv). It should be in same directory as 03_run-finemap.sh. I will fix this and push edit.

I very much appreciate your interest and realize the code requires some refactoring, which I continue to work on and would very much appreciate your feedback.

[jameskozubek]

I ran LD clumping using a random panel and my GWAS stats. Is the idea that I run LDStore only once for all a) only lead snps, or b) for all 500K snps? or c) once for each chromosomes? to create my *ld file, so should i have many *ld files, or only one, and is it only for the lead snps?
and, then, am I supposed to run Finemap, for each chromosome, for each run, should the GWAS file be limited to just the Lead snps, or all 500K snps per each chromosome?

This largely depends on resources available on your end. I chose to:

1. Determine loci from lead SNPs (see Makefile) as roughly 1 MB around each lead SNP, but merging adjacent lead SNPs prior to padding.
2. I then run LDStore on each of these loci.
3. I then run FINEMAP for each locus.

Doing it this way I was able to parallelize the process onto a HPC using a batch scheduler.

Alternatively, as you suggest, you can run LDStore on the entire genome and keep that as a single data set from which you draw SNP correlations.

By 500k SNPs, I assume you mean 500k UKB individuals. If so, I used a 50k random subset from how Bycroft et al defined the white British subset. This flag is available in the ukb_sqc_v2.txt file from UK Biobank (column in.white.ancestry.subset). I can share code with you how I did this, but was simply some awk followed by sort -R. Main reason for 50k is computational, and also using all 500k would likely yield marginal improvement.

[jameskozubek]

I see. that makes sense.

[jameskozubek]

I tried it two ways, once running LDStore per each locus + 1MB, and then again running it by combining all of these SNPs into a single file (15K snps in total in single LD store file). However, if I do it as a single file, Finemap creates a resulting "cred" file in which there are SNPs from multiple different chromosomes in the same causal set (say a set of 5 snps but one is on chr21 and another on chr1). Can this be meaningful, or is it nonsense to use Finemap in this way?

15k SNPs in total across entire genome? If so, where the SNPs filtered a priori? Normally SNPs should not be agressively filtered prior to FINEMAP. For instance, each of the 1Mbp loci we processsed for Ei contained ~1000 SNPs. It would be good to know more specifics on your analysis if you can provide. If you feel uncomfortable doing so publicly you can send me an email.

[jameskozubek]

I thought the idea was you select your top hits gwas SNPs, so say 35 snps. then you go out 1MB. Then I have 35 files created from LDStore each with about ~500 to 1000 snps. then I run finemap on each of the 35 files separately. I also tried to run finemap by combining all 35 loci + 1MB into a single LD file, which is about 15K snps.

[jameskozubek]

First: I am wondering if this makes sense, to create a single LD file which incorporates snps from different loci on different chromosomes, or if it only is appropriate to create separate LD files for each loci. Second question: How do I evaluate what is a good posterior probability? Should I only trust the cred files with > 0.95 as it says in the log file? Or, can I trust all the results in the cred files. Thank you for your time!

Got it. Apologies for my confusion. I kept everything separate. Unclear from the FINEMAP documentation whether you can merge them prior to analysis.

[jameskozubek]

I agree. It's an interesting question. I will see if the Finemap team has any opinions on this.

First: I am wondering if this makes sense, to create a single LD file which incorporates snps from different loci on different chromosomes, or if it only is appropriate to create separate LD files for each loci. Second question: How do I evaluate what is a good posterior probability? Should I only trust the cred files with > 0.95 as it says in the log file? Or, can I trust all the results in the cred files. Thank you for your time!

I would potentially make a single LDStore file simply to have a reference dataset from which you can pull the correlations from, but I would run FINEMAP per locus.

Filtering FINEMAP results is subjective to some extent. log10(BF)>2 is typically used and has shown good enrichment for functional genomic annotations and proximity to causal genes. Likely same for cred > 0.95. Of note, we removed loci that failed to converge, ie where the posterior probability of causal config containing k snps is > 0.95 at max_k.