Abstract
Gene regulatory networks (GRNs) describe how transcription factors (TFs) coordinate target gene expression to establish and maintain cellular identity. Although recent advances in single-cell transcriptomic and epigenomic technologies have greatly improved computational GRN inference, most existing approaches rely primarily on gene expression and chromatin accessibility, leaving many additional layers of gene regulation underutilized. Incorporating complementary regulatory information has the potential to improve biological accuracy while providing deeper insight into the mechanisms governing cell identity.In this project, we utilized both noncanonical and canonical (tool-supported) methods for integrating diverse data modalities into the GRN inference process to study how these data can refine the resulting network. Using neonatal mouse cochlear hair cells (HCs) as a model system, conventional CellOracle baseline networks were inferred from paired single-nucleus RNA-seq and ATAC-seq data from cochlear tissue and iteratively refined by incorporating HC-specific chromatin conformation (Micro-C), transcription factor occupancy (CUT&RUN), histone activation state (H3K27ac), and computational transcription factor footprinting (with bulk ATAC-seq). Existing and newly collected datasets were used in this process.
These data-guided refinements substantially altered network architecture by introducing novel regulatory relationships, removing unsupported interactions, and reprioritizing retained TF-target links. Refinement improved biological specificity while preserving the core regulatory hierarchy recovered by conventional GRN inference. Independent validations demonstrated reductions in false positive predictions across refinement iterations, while comparisons with SCENIC+ and LINGER showed that these refinement principles are broadly applicable across different modeling and statistical approaches.
The refined CellOracle network recapitulated established principles of cochlear HC development and maturation, identified biologically meaningful regulatory modules, and provided a practical framework for downstream analysis and application. Our result provides insights into the complexities of HC identity, including in relevant contexts such as regeneration, which can guide us toward treatment of hearing loss.
Altogether, this work highlights the importance of recognizing the many layers of biological gene regulation and seeking new ways to include them in computational modeling. In demonstrating the efficacy of our approach, we have also established that our refined neonatal cochlear HC network can serve as a reliable resource for future studies. These achievements help advance both auditory research and computational biology.