SPARK: Bayesian structure learning of enhancer-based gene regulatory networks from single-cell multiomics data and prior knowledge

Abstract

Enhancer-based gene regulatory networks (eGRNs) linking enhancers with recruited transcription factors (TFs) and downstream target genes can help decipher complex regulatory mechanisms, such as the function of GWAS variants at enhancers. Molecular assays that profile gene expression and chromatin state jointly from the same cells (such as 10x Multiome) are a rich source of input data for eGRN reconstruction. However, these assays lack mechanistic evidence for enhancer-gene and TF-enhancer links, which can be contributed by prior knowledge from other assays and analyses. Here, we present SPARK, an eGRN reconstruction framework that formally combines single-cell multiome data with mechanistic prior knowledge. SPARK represents eGRNs as directed acyclic graphs (DAGs) and learns them using Bayesian inference, with a prior on the network structure constructed from heterogeneous sources of evidence. In a simulation study, we show that SPARK benefits from prior knowledge and outperforms existing DAG-based and regression-based algorithms. Finally, we use SPARK to study the regulatory circuitry of enhancers harboring immune disease GWAS variants in human primary monocytes, revealing known and novel regulatory links as candidates for experimental validation.