Adv Sci (Weinh). 2026 Feb 15:e14446. doi: 10.1002/advs.202514446. Online ahead of print.
ABSTRACT
MicroRNAs (miRNAs) are pivotal post‑transcriptional regulators whose single‑cell behavior has remained largely inaccessible due to technical barriers in single-cell small‑RNA profiling. We present SiCmiR, a two‑layer neural network that predicts miRNA expression profiles from only 977 LINCS L1000 landmark genes, thereby reducing sensitivity to dropout in single-cell RNA-seq (scRNA-seq) data. Proof‑of‑concept analyses illustrate how SiCmiR can uncover candidate hub‑miRNAs in bulk-seq cell lines and hepatocellular carcinoma, scRNA-seq pancreatic ductal carcinoma, and ACTH‑secreting pituitary adenoma and extracellular vesicle (EV)‑mediated crosstalk in glioblastoma. Trained on 6,462 TCGA paired miRNA-mRNA samples, SiCmiR attains state‑of‑the‑art accuracy on cancers and generalizes to unseen cancer types and drug perturbations. We next construct SiCmiR‑Atlas, containing 362 public datasets, 9.36 million cells, and 726 cell types, which is the first dedicated database of single‑cell mature miRNA expression, providing interactive visualization, biomarker identification, and cell‑type‑resolved miRNA-target networks. SiCmiR transforms bulk‑derived statistical power into a single‑cell view of miRNA biology and provides a community resource for biomarker discovery. SiCmiR Atlas is available at https://awi.cuhk.edu.cn/∼SiCmiR/.
PMID:41691474 | DOI:10.1002/advs.202514446
