Serum-MiR-CanPred: deep learning framework for pan-cancer classification and miRNA-targeted drug discovery.

Cancer diagnosis at an early stage is crucial for improving overall health outcomes. However, existing cancer diagnostic techniques are mostly invasive and tend to identify the disease only in its advanced stages. MicroRNAs (miRNAs), which are small non-coding RNAs involved in gene expression regulation, are stable in serum as circulating miRNAs and have potential as non-invasive biomarkers. However, their application in pan-cancer diagnostics and therapeutics is still largely unexplored. We developed Serum-MiR-CanPred, a deep learning framework using a multi-layer perceptron (MLP) trained on serum miRNA expression data from 20,271 samples across 12 cancer types and healthy controls from GEO databases. The model achieves robust pan-cancer classification (AUC = 96.87%, accuracy = 96%) with a consensus set of 88 miRNAs. Validation using external datasets demonstrated its generalizability and clinical potential. SHapley Additive exPlanations (SHAP) identified hsa-miR-5100 as a key biomarker, dysregulated in cancers including lung, bladder, and gastric carcinomas. Pathway analysis linked these miRNAs to cancer-related processes like VEGFA-VEGFR2 signalling. Molecular docking of pre-mir-5100 with rDock, identified AC1MMYR2 as a potential high-affinity ligand, with binding stability confirmed by molecular dynamics simulations using GROMACS In conclusion, Serum-MiR-CanPred integrates explainable AI with molecular modelling, advancing miRNA-based diagnostics and drug discovery for precision oncology.