miDGD: a multi-modal deep generative model predicts miRNA expression from bulk or single-cell mRNA expression
MicroRNAs (miRNAs) are key post-transcriptional regulators, yet standard bulk and single-cell RNA-seq do not capture them, leaving this regulatory layer invisible in most transcriptomic data. We present miDGD, a deep generative decoder that jointly models paired mRNA and miRNA profiles through a shared latent representation, enabling miRNA expression to be predicted from mRNA alone. Trained on tumors (TCGA), healthy tissues (GTEx), and human cell lines, miDGD recovers hundreds of miRNAs in held-out tumors (mean Spearman {rho} = 0.56), captures both tissue-specific and ubiquitous miRNAs, and preserves known miRNA--target repression and host-gene co-expression. Without label supervision, its latent space separates 32 cancer types (80% accuracy). Predictions remain stable at single-cell-like sparsity and transfer across datasets--from tumors to healthy tissues and from bulk to single cells--where miDGD outperforms existing supervised and activity-inference methods. miDGD thus unlocks miRNA regulation in the vast body of existing mRNA-only data, including single-cell datasets.