Learning sequence-based regulatory dynamics in single-cell genomics

Abstract

Epigenomics assays, such as chromatin accessibility, can identify DNA-sequence-specific regulatory factors. Models that predict read counts from sequence features can explain cell-based readouts using specific DNA patterns (genomic motifs) but do not encode the changes in genomic regulation over time, which is crucial for understanding biological events during cell transitions. To bridge this gap, we present muBind, a deep learning model that accurately predicts genomic counts of single-cell datasets based on DNA sequence features, their cell-based activities, and cell relationships (graphs) in a single architecture, enhancing the interpretability of cell transitions due to the possibility of inspecting motif activities weighted by nearest neighbors. MuBind shows competitive performance in bulk and single-cell genomics. When complemented with graphs learned from RNA-based dynamical models used as injected priors in our model, muBind enhances through motif-graph interactions the identification of transcriptional regulators explaining cell transition events, including Sox9 in pancreatic endocrinogenesis scATAC-seq, and Gli3/Prdm16 in mouse neurogenesis and human organoids scRNA-seq, both supported by independent evidence, including associations between chromatin and motif activities over pseudotime, TF-gene expression patterns, and biological knowledge of these regulators. muBind advances our understanding of cell transitions by revealing regulatory motifs and their interactions, providing valuable insights for genomic research and gene regulatory network dynamics. It is available at https://github.com/theislab/mubind.