Benchmarking Machine Learning Methods for Multi-Omics Data Integration in Cancer Subtyping

The integration of multi-omics data through machine learning has become a pivotal approach for enhancing cancer subtyping, offering deeper insights into tumor heterogeneity and supporting the advancement of personalized medicine. Despite the growing adoption of various integration methods, the lack of standardized evaluation frameworks poses a significant challenge, complicating the assessment and comparison of their effectiveness across different cancer types and data configurations. To address this issue, we constructed comprehensive benchmarking datasets using data from The Cancer Genome Atlas (TCGA), encompassing nine distinct cancer types and exploring all eleven possible combinations of four key multi-omics data types: genomics, transcriptomics, proteomics, and epigenomics. In this study, we performed an in-depth evaluation of twelve well-established machine learning methods for multi-omics integration, focusing on critical performance metrics such as clustering accuracy, clinical relevance, robustness, and computational efficiency. Our results revealed that iClusterBayes achieved an impressive silhouette score of 0.89 at its optimal k, followed closely by Subtype-GAN (0.87) and SNF (0.86), indicating their strong clustering capabilities. Notably, NEMO and PINS demonstrated the highest clinical significance, with log-rank p-values of 0.78 and 0.79, respectively, effectively identifying meaningful cancer subtypes. Furthermore, NEMO ranked highest overall with a composite score of 0.89, showcasing its strengths in both clustering and clinical metrics, thus establishing it as a leading method in multi-omics integration. In robustness testing, LRAcluster emerged as the most resilient method, maintaining an average normalized mutual information (NMI) score of 0.89 even as noise levels increased, which is crucial for real-world data applications. In terms of computational efficiency, Subtype-GAN stood out as the fastest method, completing analyses in just 60 seconds, while NEMO and SNF demonstrated commendable efficiency with execution times of 80 and 100 seconds, respectively. Interestingly, our findings highlighted that more data does not always equate to better outcomes; in fact, using combinations of two or three omics types frequently outperformed configurations that included four or more types due to the introduction of increased noise and redundancy. These insights are particularly relevant for researchers and practitioners in precision oncology, as they provide a framework for developing more effective, data-driven strategies tailored to individual tumor profiles. By emphasizing the context-dependent nature of multi-omics data integration, this study lays the groundwork for future advancements in cancer research and personalized treatment approaches, ultimately enhancing the efficacy of precision medicine and improving patient outcomes. As the landscape of multi-omics continues to evolve, our research underscores the importance of refining integration methods and selecting appropriate data types, ensuring that clinicians can effectively leverage these advanced tools in their quest to optimize cancer treatment strategies.