Abstract
As a highly heterogeneous and complex disease, the identification of cancer's molecular subtypes is crucial for accurate diagnosis and personalized treatment. The integration of multi-omics data enables a comprehensive interpretation of the molecular characteristics of cancer at various biological levels. In recent years, an increasing number of multi-omics clustering algorithms for cancer molecular subtyping have been proposed. However, the absence of a definitive gold standard makes it challenging to evaluate and compare these methods effectively. In this study, we developed a general framework for the comprehensive evaluation of multi-omics clustering algorithms and introduced an innovative metric, the accuracy-weighted average index, which simultaneously considers both clustering performance and clinical relevance. Using this framework, we performed a thorough evaluation and comparison of 11 state-of-the-art multi-omics clustering algorithms, including deep learning-based methods. By integrating the accuracy-weighted average index with computational efficiency, our analysis reveals that PIntMF demonstrates the best overall performance, making it a promising tool for molecular subtyping across a wide range of cancers.