Causal network deconvolution

The project is sub-divided into two parts:

1. Simulations

2. Application to real data

3. Simulation study

Questions to answer in this section: • How bad is confounding for ND-correlation matrix? • How bad is pleiotropy for ND-MR matrix? • Sample size differences between ND-cor and ND-MR? • How are cycles handled between the two methods? • Missing nodes impact on ND-cor vs ND-MR • Sparsity impact on ND-cor vs ND-MR • Measurement error impact on ND-cor vs ND-MR

Decisions to make for analysis • Number of iterations - look at the standard error of the points in the existing results - if it is very high then need more iterations • Number of nodes range: 10, 100, 1000? • Sample size range: 100 or 5000 • Confounding level: 0-100 • Pleiotropy level: 0-100 • Consider higher order confounding later on - e.g. confounding by cell type in bulk tissue analysis of gene expression levels • Does the intersection between ND-cor and ND-MR improve AUC?

Simulations to evaluate how to deconvolve the matrix of indirect causal effects into one of direct causal effects. Using network deconvolution method.

To do

• Compare against ND of observational correlation matrix
• Use ND normalisation method, test on different levels of sparseness
• Evaluate performance with
  • cycles in the graph
  • measurement error
  • unmeasured confounding / incomplete information
• Evaluate sensitivity to biases in the MR matrix
• Compare to multivariable MR

Setup

Create a config.json file with relevant paths:

{
    "genodir": "",
    "phendir": "",
    "igddir": ""
    "pipelinedir": "",
    "datadir": "",
    "resultsdir": "",
}

• genodir = path to ukbb bgen files
• phendir = path to phesant formatted phenotype files
• igddir = path to GWAS VCF files for ukb-b batch
• pipelinedir = path to original ukb-b gwas pipeline dir
• datadir = where to store generated data
• resultsdir = where to store generated results

To run everything on bc4

Use Snakemake to orchestrate the analysis. This will submit the jobs to slurm. So when you're in a screen session, to run the snakemake process:

module add languages/anaconda3/5.2.0-tflow-1.11
source ~/.bash_profile
snakemake -prk \
-j 100 \
--cluster-config bc4-cluster.json \
--cluster "sbatch \
  --job-name={cluster.name} \
  --partition={cluster.partition} \
  --nodes={cluster.nodes} \
  --ntasks-per-node={cluster.ntask} \
  --cpus-per-task={cluster.ncpu} \
  --time={cluster.time} \
  --mem={cluster.mem} \
  --output={cluster.output}"