Systematic comparison of colocalization methods using protein quantitative trait loci

Colocalization is frequently performed as a step to triage findings from genetic investigations linking molecular and disease data. However, the reliability and consistency of the various colocalization methods is not well understood. In particular, it is unclear whether a non-colocalization result should be taken as a definitive sign that two traits do not share a genetic cause in the region of interest, or merely as suggestive evidence. We use protein QTL data to benchmark four colocalization methods (coloc, coloc-SuSiE, prop-coloc, and colocPropTest), considering whether the methods conclude there is colocalization between datasets representing associations with the same protein. We consider a baseline scenario in which the associations come from the same dataset split in half at random, and scenarios in which the associations are with the same protein, but measured on different platforms or in different populations. In the baseline scenario, all methods report colocalization for the majority of proteins. In other scenarios, methods do not consistently report colocalization, they often report non-colocalization, and they often disagree. In the worst-case scenario, colocalization was only agreed by all four methods for 20% of proteins, despite our experiment being constructed to select for cases where colocalization is expected. This suggests caution in performing and interpreting colocalization analyses is warranted.