miRSiC: a regulatory-aware machine learning framework for microRNA expression inference across bulk and single-cell transcriptomes.

MicroRNAs (miRNAs) are key post-transcriptional regulators embedded in gene regulatory networks between upstream transcription factors (TFs) and downstream target genes (TGs), yet most computational approaches infer miRNA expression using unstructured transcriptomic features without explicitly modeling their regulatory architecture. In this study, we develop miRSiC, an interpretable machine learning framework that integrates TFs and experimentally validated TGs for regulatory-aware miRNA expression inference. miRSiC formulates prediction as miRNA-specific regression tasks and evaluates three regulatory configurations (TF-only, TG-only, and combined TF-TG) using Light Gradient Boosting Machine. Applied to The Cancer Genome Atlas (TCGA) breast cancer cohort, the integrated model achieves superior performance (mean Spearman correlation = 0.5462 across 326 miRNAs), outperforming single-layer models and demonstrating the effectiveness of incorporating both upstream and downstream regulatory signals. Feature importance analysis and regulatory network reconstruction show that selected features are enriched in biologically coherent TF-miRNA-target circuits. Prediction performance varies across breast cancer subtypes, reflecting differences in regulatory patterns and sample size. Cross-platform evaluation further reveals that models trained on bulk transcriptomes do not generalize to single-cell data due to distributional shifts and sparsity; however, domain-specific retraining partially restores performance. Together, miRSiC provides an interpretable and biologically grounded framework for miRNA expression inference, highlighting the importance of modeling regulatory context and adopting domain-aware strategies across bulk and single-cell transcriptomic data.