Ovarian cancer (OV) remains the deadliest gynecological malignancy, with non-coding RNA-mediated transcriptomic deregulation significantly influencing its prognosis and heterogeneous progression. In this study, we prioritized miRNA-mediated gene expression profiles by identifying key negative correlations between miRNA-mRNA pairs. We developed a machine learning-based non-coding index (NCI), incorporating a four-gene signature (GAS1, GFPT2, ZFHX4, and KCNA1) to predict patient prognosis and therapeutic response. Validation across multiple datasets revealed that OV patients with higher NCI scores had significantly poorer survival outcomes and resistance to immunotherapy. Additionally, we established a four-class subtyping taxonomy through unsupervised clustering, validated in four independent datasets. The S1 and S3 subtypes were characterized by high NCI scores, abundant stromal and immune infiltration, with the S3 subtype exhibiting the worst survival. Conversely, the S2 subtype showed downregulation of immune response genes, while the S4 subtype displayed epithelial differentiation and favourable prognosis. Integrative analyses of bulk and single-cell transcriptomic data revealed that the S3 subtype had a significantly higher fibroblast proportion compared to other subtypes, whereas the S1 subtype was marked by high T cell content. Through ridge regression-based drug sensitivity analyses, we prioritized candidate therapeutics for each subtype. Notably, the S3 subtype demonstrated sensitivity to dasatinib but resistance to methotrexate. Finally, we developed a user-friendly Shiny-based website to facilitate the application of our prognostic and subtype classification models (https://jli-bioinfo.shinyapps.io/NCI_online/). This study establishes a critical prognostic marker and proposes a novel molecular classification framework grounded in miRNA-regulated gene expression profiles, advancing our understanding of the non-coding mechanisms driving OV heterogeneity.