Genome-wide regulatory model from MPRA data predicts functional regions, eQTLs, and GWAS hits

Abstract

Massively-parallel reporter assays (MPRA) enable unprecedented opportunities to test for regulatory activity of thousands of regulatory sequences. However, MPRA only assay a subset of the genome thus limiting their applicability for genome-wide functional annotations. To overcome this limitation, we have used existing MPRA datasets to train a machine learning model that uses DNA sequence information, regulatory motif annotations, evolutionary conservation, and epigenomic information to predict genomic regions that show enhancer activity when tested in MPRA assays. We used the resulting model to generate global predictions of regulatory activity at single-nucleotide resolution across 14 million common variants. We find that genetic variants with stronger predicted regulatory activity show significantly lower minor allele frequency, indicative of evolutionary selection within the human population. They also show higher over-lap with eQTL annotations across multiple tissues relative to the background SNPs, indicating that their perturbations in vivo more frequently result in changes in gene expression. In addition, they are more frequently associated with trait-associated SNPs from genome-wide association studies (GWAS), enabling us to prioritize genetic variants that are more likely to be causal based on their predicted regulatory activity. Lastly, we use our model to compare MPRA inferences across cell types and platforms and to prioritize the assays most predictive of MPRA assay results, including cell-dependent DNase hypersensitivity sites and transcription factors known to be active in the tested cell types. Our results indicate that high-throughput testing of thousands of putative regions, coupled with regulatory predictions across millions of sites, presents a powerful strategy for systematic annotation of genomic regions and genetic variants.