Epigenetic Combinatorial Patterns Predict Disease Variants.

Most genetic variants identified in genome-wide association studies are noncoding and are likely tagging nearby causal variants. It is a challenging task to pinpoint the precise locations of disease-causal variants and understand their functions in disease. A promising approach to improve fine mapping is to integrate the functional data currently available on hundreds of human tissues and cell types. Although there are several methods that use functional data to prioritize disease variants, they mainly use linear models, or equivalent naive likelihood-based models for prediction. Here, we investigate whether study of the combinatorial patterns of functional data across cell types can improve prediction accuracy for disease variants. Using functional annotation in 127 human cell types, we first introduce a Bayesian method to identify recurring cell-type-specificity partitions on the scale of the genome. We show that our de novo identification of epigenome partition patterns agrees well with known cell-type origins and that the associated functional elements are strongly enriched in disease variants. Using epigenetic cell-type specificity in addition to enrichment of functional elements, we further demonstrate that the power to predict disease variants can be greatly improved over that achievable with linear models. Our approach thus provides a new way to prioritize disease functional variants for testing.