In kevinlkx/Mapgen: Multi-function software that performs enrichment, functionally-informed genetic fine-mapping and gene mapping

knitr::opts_chunk$set(echo = TRUE, eval=TRUE)

In this tutorial, we perform fine-mapping (using SuSiE) with functional priors computed from the enrichment analysis using TORUS.

Load the packages.

library(mapgen)

Input data

Please see the data preparation tutorial about preparing GWAS summary statistics.

We need a bigSNP object for the reference genotype panel, which will be used for computing LD matrices. See the data preparation tutorial for details.

SNP-level priors

Let's load the processed GWAS summary statistics obtained from the data preparation tutorial and the SNP-level priors using run_torus() obtained from the enrichment analysis tutorial.

gwas.sumstats <- readRDS(system.file('extdata', 'test.processed.sumstats.rds', package='mapgen'))
torus.res <- readRDS(system.file('extdata', 'test.torus.res.rds', package='mapgen'))

We can add the SNP-level priors (column called 'torus_prior') to our summary statistics.

sumstats.for.susie <- prepare_susie_data_with_torus_result(sumstats = gwas.sumstats, 
                                                           torus_prior = torus.res$snp_prior)

We could set filter == "pval" to limit loci with GWAS p-value cutoff (default: 5e-8), or set filter == "torus_fdr" by the FDR cutoff (default: 0.1) from running run_torus() with option = "fdr".

Fine-mapping

We can perform fine-mapping with SNP-level priors using SuSiE, you will need the susieR package for the steps below.

library(susieR)

To use SNP-level priors computed from TORUS, we use run_finemapping() with priortype = 'torus'.

You can also run fine-mapping using uniform prior (without SNP-level priors) with priortype = 'uniform'. (You don't need to run TORUS if you run fine-mapping using uniform prior.)

You can use the bigSNP object prepared earlier in the data preparation tutorial. If you provide bigSNP instead of LD matrices, it will compute R matrices using the genotype data in the bigSNP object.

# you will need the bigsnpr package if you use the bigSNP object.
library(bigsnpr)
susie.res <- run_finemapping(sumstats = sumstats.for.susie, 
                             bigSNP = bigSNP, 
                             priortype = 'torus', 
                             L = 1)

We did not provide sample size (n) for this toy example, but it is highly encouraged to provide that (set argument n = sample size) in the real data analysis.

If you have the LD matrices, you can use those instead of the bigSNP object.

We have pre-computed the LD matrices of European samples from UK Biobank. They can be downloaded here.

See Fine-mapping with UK Biobank LD tutorial and [LD diagnosis tutorial][ld-diagnosis-tutorial] for details.

We ran SuSiE with L = 1 here, which allows a single causal signal for each LD block and is robust to mismatching LD patterns.

susie.res is a list of SuSiE results, one for each locus. We can merge GWAS summary statistics with SuSiE result:

finemap.sumstats <- merge_susie_sumstats(susie.res, sumstats.for.susie)

We now have a new column called susie_pip, which is the probability of a SNP being causal estimated using SuSiE.

Note: Inconsistencies between GWAS z-scores and the LD reference could potentially inflate PIPs from SuSiE finemapping. It would be helpful to check potential LD mismatch issue, for example using DENTIST or the diagnostic from susie_rss.

See Fine-mapping with UK Biobank LD tutorial for details.

So to be conservative, you could use L = 1 if you have "out-of-sample" LD reference, and focus on the loci with significant GWAS signals (e.g. pval < 5e-8).

kevinlkx/Mapgen documentation built on April 2, 2024, 5:48 a.m.