Reli-miR: A Reliability-Aware Framework for Virtual miRNA Reconstruction from mRNA Expression with Downstream Preservation Assessment

Background. MicroRNAs (miRNAs) are central post-transcriptional regulators in cancer research. Many cohorts provide mRNA profiles without matched miRNA measurements, and not all miRNAs are equally reconstructable from mRNA. Using all outputs indiscriminately compromises downstream analyses. Methods. Reli-miR is a reliability-aware framework for virtual miRNA reconstruction from mRNA. It uses a strict train/validation/test split, compares six models and two baselines under a predefined validation rule, assesses stability across 20 resampled patient splits, estimates miRNA- and sample-level reliability, and evaluates six downstream dimensions against equal-size random subsets across eight cohorts. Results. On 366 paired samples (3000 mRNAs, 846 miRNAs), PCA_ElasticNet was the formal best model (validation r = 0.4705); ElasticNet led on the test set (r = 0.4498), and leading models overlapped. miRNA-level reliability strongly calibrated independent test error (Spearman = −0.5978), with the association negative in all 20 resplits (medians −0.56 to −0.57). At 0.7, 33 retained miRNAs reached sample correlation 0.8442 and Cox logHR preservation 0.8953, outperforming equal-size random subsets in every split. Across eight cohorts, reliability retained its ranking action, and the high-reliability subset outperformed each cohort, although mean R2 was negative in all eight. Conclusions: Reliability-aware filtering identifies trustworthy miRNA subsets, with leading linear models performing comparably.