Abstract
Objective: Accurate cancer prognosis prediction is essential for personalized treatment planning and clinical decision-making. With the rapid development of high-throughput sequencing technologies, multi-omics data such as messenger RNA expression, microRNA profiles, and DNA methylation provide complementary molecular perspectives on tumor progression. However, the high dimensionality, heterogeneity, interpretability, and complex cross-modal dependencies of multi-omics data pose substantial challenges to existing survival prediction models. Methods: To address these challenges, we propose an adaptive Omics-to-Image Transformer framework for Cancer prognosis Evaluation (OTCE), which converts heterogeneous multi-omics data into unified pseudo-image representations and further integrates them into multichannel image representations, facilitating effective cross-modal feature learning. A parallel multiview deep neural network composed of multiple functional modules is designed to capture global distributional characteristics, local spatial patterns, and long-range cross-modal dependencies, respectively. Results: Extensive experiments on 6 cancer datasets demonstrate that OTCE consistently outperforms state-of-the-art traditional and deep-learning-based survival models, achieving an average improvement of 7.7% in concordance index (C-index). Furthermore, by integrating Shapley-additive-explanations-based feature attribution with differential expression analysis, OTCE identifies 7 prognostic candidate biomarkers in kidney renal clear cell carcinoma, whose biological relevance is further supported by single-cell and spatial transcriptomic analyses. Conclusion: These results indicate that OTCE effectively improves the accuracy and robustness of multi-omics cancer prognosis prediction while enhancing model interpretability. The proposed framework provides a scalable and biologically meaningful solution for integrative survival analysis and offers valuable insights for prognostic biomarker discovery across multiple cancer types. The code of OTCE is available at https://github.com/fsct135/fsct135-2/tree/main.