Inferring transcriptional activation and repression activity maps in single-nucleotide resolution using deep-learning

Abstract Recent computational methods for inferring cell type-specific functional regulatory elements have used sequence and epigenetic data. Active regulatory elements are characterized by open-chromatin state, and the novel experimental technique ATAC-STARR-seq couples ATAC-seq assays, which capture such genomic regions, with a functional assay (STARR-seq) to selectively examine the regulatory activity of accessible DNA. ATAC-STARR-seq may thus provide data that could improve the quality of computational inference of active enhancers and silencers. Here, we propose a novel regression-based deep learning (DL) model that utilizes such data for predicting single nucleotide activation and repression maps. We found that while models using only sequence and epigenetics data predict active enhancers with high accuracy, they generally perform poorly in predicting active silencers. In contrast, models building also on data of experimentally identified enhancers and silencers do substantially better in the identification of active silencers. Our model predicts many novel enhancers and silencers in the model lymphoblastoid cell line GM12878. Epigenetic signatures of the novel regulatory elements detected by our model resemble the ones shown by the experimentally validated enhancers and silencers in this cell line. ChIP-seq enrichment analysis in predicted novel silencers identify a few significant enriched transcriptional repressors such as SUZ12 and EZH2, which compose the PRC2 repressive complex. Intersection with GWAS data found that the novel predicted enhancers are specifically enriched for risk SNPs of the Lupus autoimmune disease. Overall, while silencers are still poorly understood, our results show that our DL-model can be used to complement the experimental results on regulatory element discovery.