Cell type–specific functions of nucleic acid-binding proteins revealed by deep learning on co-expression networks

Nucleic acid-binding proteins (NABPs) exhibit cell type–specific regulatory functions, but their target genes and biological roles remain incompletely characterized due to the limitations of current experimental approaches. Here, we present a deep learning framework that integrates gene co-expression correlations to predict NABP regulatory targets and infer their functions across diverse cellular contexts, without requiring binding site or motif information. Substituting low-informative input features with co-expression-derived interactions improved gene expression prediction accuracy. Predicted targets showed strong concordance with ChIP-seq and eCLIP binding sites, and this agreement was significantly greater than for randomly selected gene sets. Functional enrichment and ChatGPT-assisted inference revealed biologically meaningful annotations, including cell type–specific functions such as circadian regulation by AKAP8 in cancer cells and glycolytic control by PKM. Collectively, this integrative framework—combining deep learning, co-expression networks, and large language models—enables the systematic discovery of both known and previously uncharacterized NABP functions in a cell type–specific manner.