A brief outline of colocalisation analysis

The coloc package can be used to perform genetic colocalisation analysis of two potentially related phenotypes, to ask whether they share common genetic causal variant(s) in a given region. There are a few key references which this vignette will not duplicate (see below).

In brief, two approaches can be implemented. The proportional testing approach 12 has now been moved to its own package, coloc.prop.

This package implements the more population enumeration approach.

You can read about how to prepare your data in this vignette or read the vignettes listed in one of the sections below to understand how coloc works.

A single causal variant assumption

Claudia Giambartolomei and Vincent Plagnol proposed the enumeration method, which makes use of Jon Wakefield’s work on determining approximate Bayes Factors from p values 3 to generate a colocalisation analysis 4, implemented in the function coloc.abf(). By assuming there is at most one causal variant per trait, every possible configuration can be individually enumerated and evaluated, and aggregating over these allows us to gauge the relative support for models which support colocalisation to those that don’t.

You can see more about the enumeration approach on this blogpost.

See vignette: enumeration

Sensitivity analysis

As a Bayesian method, coloc.abf() requires the user to specify prior probabilities of SSNP causality and colocalisation. Post-hoc sensitivity analysis can be used to assess whether results are robust across a range of plausible priors.

See vignette: sensitivity

Deprecated: relaxing the single causal variant assumption through conditioning

The single variant assumption can be relaxed through conditioning. We have implemented a conditioning step within coloc, which we hope will increase use of conditioning, and proposed an alternative, masking.

See vignette: conditioning/masking

An improved approach to relaxing the single causal variant assumption: SuSiE

The sum of single effects regression method proposed by Wang et al 6 can instead simultaneously decompose multiple signals from the marginal summary stats, and appears to work better than conditioning.

See vignette: SuSiE