A Context-Aware Graph Transformer Framework for microRNA-Gene Regulatory Inference Across Bulk Tumor and Single-Cell Cancer Data.

Background: MicroRNAs are key post-transcriptional regulators of gene expression and contribute to cancer progression, tumor heterogeneity, and context-dependent regulatory rewiring. However, most computational approaches rely on sequence-based target prediction or bulk expression association and are not designed to jointly model regulatory priors, expression context, and heterogeneous cancer states, particularly when matched single-cell microRNA/mRNA co-profiling data are scarce. Methods: We developed a context-aware graph transformer framework for microRNA-gene regulatory analysis across biological resolutions. The framework represents microRNAs, genes, and biological contexts as a heterogeneous graph, where contexts correspond to individual cells in single-cell data and tumor samples or subtype-defined profiles in bulk cohorts. Heterogeneous graph transformer learning generated regulatory embeddings, Bayesian topology optimization refined candidate microRNA-gene interactions, and a dominance-based competition layer with Dominance Share scoring identified master regulators and cooperative target modules. Results: We applied miR-CellMap to matched single-cell miRNA/mRNA co-sequencing data from K562 leukemia cells and paired bulk cancer datasets spanning pan-cancer and subtype-specific cohorts, including breast, colon, glioblastoma, lower-grade glioma, and ovarian cancer. The framework identified recurrent and dataset-specific miRNA regulatory programs, including regulators such as miR-186-5p, miR-214-3p, miR-27a-3p, and let-7 family members. Embedding-derived context analysis showed that predicted miRNA target programs were consistently closer to observed context-specific gene programs than random matched gene programs across all seven datasets. Dominance Share analysis further identified cooperative target modules and co-repressed target programs, supporting the use of miR-CellMap for interpretable cancer-focused miRNA regulatory discovery. Conclusions: This framework provides an interpretable strategy for mapping conserved, cancer-specific, and context-dependent microRNA-gene regulatory programs across single-cell and bulk cancer datasets.