Precision medicine leverages multi-omics data to enhance cancer subtype classification, yet significant challenges remain in integrating heterogeneous data types, handling sparse datasets, and modeling complex biological interactions. Traditional approaches often struggle to dynamically capture cross-omics relationships, limiting their predictive accuracy and clinical applicability.To overcome these limitations, we propose DeepOmicFusion, an attention-based deep learning framework that integrates miRNA expression, copy number variation (CNV), and miRNA–target interaction networks for accurate multiclass cancer subtype classification. Ablation studies demonstrate that multimodal integration improves classification performance by approximately 4.3% compared to single-omics approaches. Comparative evaluations against conventional machine learning models confirm improved robustness and generalizability across TCGA and GDC datasets. The attention-based fusion mechanism prioritizes biologically relevant interactions, while transfer learning enables effective knowledge transfer from pan-cancer to cancer-specific contexts. Overall, DeepOmicFusion provides a scalable and interpretable solution for multi-omics cancer subtype classification and shows strong potential for advancing precision oncology.