Objective: The integration of multi-omics data to uncover the biological mechanisms of human diseases remains a significant challenge in bioinformatics. While deep learning (DL) has emerged as a powerful tool for this task, current methods often fail to model the complex correlations among features and samples, limiting both predictive performance and interpretability.

Methods: To address this, we propose MOFRCDLANet (Multi-Omics Feature Reordering Correlations Deep Attention Network), a novel framework for predicting tumor recurrence and identifying biomarkers. Our model introduces a feature reordering strategy to prioritize prognostically relevant features. It then employs a self-attention module coupled with Maximum Mean Discrepancy (MMD) and contrastive regularization to learn robust latent representations that capture cross-sample relationships and align feature distributions across omics types. Finally, an attribution-based method identifies key biomarkers, providing biological insight into the model's predictions.

Results: Extensive experiments on ten TCGA cancer datasets demonstrate that MOFRCDLANet outperforms state-of-the-art methods across key metrics, including accuracy and AUC. The top genes identified by the model were biologically validated through enrichment analyses (KEGG and GO), confirming their relevance to cancer pathways and reinforcing the framework's efficacy.

Conclusion: MOFRCDLANet provides a robust, interpretable solution for multi-omics integration, advancing both the predictive accuracy and mechanistic understanding of cancer progression. This work offers a valuable tool for precision oncology, enabling improved prognostic stratification and biomarker discovery.