Estimating HLA disease associations using similarity trees

Abstract

The human leukocyte antigen (HLA) is associated with many (infectious) disease outcomes. These associations are perhaps best documented for HIV-1. For example, the HLA-B*58:01 allele is associated with control of the virus, while HLA-B*18:01 is considered detrimental. In HLA disease association studies, it is often ignored that certain HLA molecules are functionally very similar to others. For instance, HLA-B*18:03 differs \"only\" at 3 positions in its peptide binding site from HLA-B*18:01, and not surprisingly, HLA-B*18:03 is also associated with fast progression to AIDS. Here, we present a Bayesian method that takes functional HLA similarities into account to find HLA associations with quantitative traits such as HIV-1 viral load. The method is based on the so-called phylogenetic mixed model (a model for the evolution of a quantitative trait on the branches of a phylogeny), and can easily be modified to study a wide range of research questions, like the role of the heterozygote advantage, or KIR ligands on disease outcomes. We show that in the case of HIV-1, our model is significantly better at predicting set-point virus load than a model that ignores HLA similarities altogether. Furthermore, our method provides a comprehensible visualization of HLA associations. The software is available online at www.github.com/chvandorp/MHCshrubs