Widespread Non-Additive and Interaction Effects Within Human Leukocyte Antigen Loci Modulate the Risk of Autoimmune Diseases

Abstract

Autoimmune diseases occur when the immune system mistakenly targets normal tissues. These diseases comprise a wide range of conditions, such as systemic lupus erythematosus, inflammatory bowel disease, and multiple sclerosis. For most autoimmune diseases, the strongest genetic risk is conferred by the human leukocyte antigen (HLA) genes, which are located within the major histocompatibility complex (MHC). The proteins encoded by HLA genes present peptide antigens to T cells, which then determine if a given antigen represents a self-peptide or a foreign peptide. In certain cases, this process can go awry, leading to T cell-mediated destruction of healthy tissues. Importantly, each HLA molecule can present a different set of peptide antigens, and patients who are heterozygous at a certain HLA gene possess a larger antigen-binding repertoire.

Currently, most models that investigate the relationship between HLA alleles and autoimmunity assume that genetic variants combine on a log-additive scale (i.e. the logarithm of the odds ratio is additive). Under a log-additive model, the log odds for a patient with one copy of a risk allele (heterozygous) is exactly half that of a patient with two copies of the allele (homozygous). However, because patients with heterozygous HLA genotypes can present a broader array of antigens, they may be more likely to present a pathogenic autoantigen that leads to autoimmunity. Therefore, I hypothesized that a non-additive model accounting for heterozygous effects among HLA alleles might yield a more accurate estimate of autoimmune disease risk.

To test this hypothesis, I examined a total of 25,835 patients across five autoimmune conditions: rheumatoid arthritis (RA), type 1 diabetes (T1D), psoriasis, achalasia, and celiac disease. I used an imputation algorithm developed by our lab to infer classical HLA alleles from dense genotype data across the MHC region. In four out of five diseases, I observed that accounting for non-additive effects significantly improved the fit of a logistic regression model (RA, P = 2.5 × 10-12; T1D, P = 2.4 × 10-10; psoriasis, P = 5.9 × 10-6; celiac disease, P = 1.2 × 10-87). In three of these diseases, the models were further improved after accounting for interactions between specific HLA alleles (RA, P = 1.8 × 10-3; T1D, P = 8.6 × 10-27; celiac disease, P = 6.0 × 10-100). These allelic interactions generally increased disease risk for heterozygous patients, and they explained moderate but significant fractions of phenotypic variance (1.4% in RA, 4.0% in T1D, and 4.1% in celiac disease) beyond an additive model. These novel findings substantially revise our understanding of the genetic architecture of autoimmunity and may provide insight into the pathogenic mechanism of these diseases.