Integrated bioinformatics and deep learning (MLP) approach reveals a novel five miRNA prognostic signature in uveal...

Uveal melanoma (UM) is the most common primary intraocular malignancy in adults, characterized by a high rate of liver metastasis and limited systemic treatment options, highlighting the urgent need for reliable prognostic biomarkers. In this study, we applied an integrated analytic pipeline combining univariate Cox proportional hazards regression with a simple deep learning model to identify prognostic microRNAs (miRNAs) in UM. Using miRNA-seq and clinical data from 80 patients in the TCGA-UVM cohort, we first identified 60 miRNAs significantly associated with overall survival after applying Benjamini-Hochberg correction (FDR < 0.01). In parallel, we used weighted Pearson correlation to select the top stage-associated miRNAs as input features for a multilayer perceptron (MLP) neural network trained to classify tumors into early (Stage II) and late (Stage III-IV) stages. The MLP achieved moderate classification performance (AUC = 0.71), and its normalized variable importance identified 20 miRNAs strongly linked to disease progression. By intersecting the top 15 miRNAs from the neural network with those from the Cox model, we identified five candidate prognostic miRNAs: miR-4435, miR-3186, miR-1250, miR-6845, and miR-4736. Previous research has shown that miR-4435 acts as an oncogene by targeting TIMP3 in colorectal cancer, miR-3186 and miR-1250 function as tumor suppressors by regulating FOXK1 and the MAPK1/WDR1 axis, respectively, miR-6845 inhibits tumor growth through suppression of USP22 and ADGRD1, and miR-4736 modulates tumor progression via repression of CSE1L and ATG7 in colorectal and pancreatic cancers. To validate their relevance, we used the GEO2R tool to assess differential expression of these miRNAs across independent UM datasets in GEO, confirming dysregulation of candidate miRNAs. Pan-cancer analysis using dbDEMC and STARBASE revealed tissue-specific expression patterns and significant prognostic associations across multiple tumor types. Furthermore, drug-miRNA interaction mapping via the ncRNADrug database highlighted associations between our candidate miRNAs and sensitivity or resistance to several chemotherapeutics, supporting their translational potential. In conclusion, our integrative bioinformatics and AI-assisted approach identified a robust five-miRNA signature with prognostic and therapeutic implications in UM. These findings warrant further experimental validation and may support the development of miRNA-guided precision oncology strategies in uveal melanoma.