Decoupled contrastive multi-view clustering with adaptive false negative elimination for cancer subtyping

Cancer’s heterogeneity necessitates precise subtype identification for effective diagnosis and treatment, which can be achieved by integrating multi-omics data to reveal distinct molecular characteristics and enable personalized therapies. Recently, significant efforts have been made through contrastive clustering methods to efficiently identify cancer subtypes. However, existing approaches remain limited in effectively capturing inter- and intra-view relationships in multi-omics data. Additionally, most cancer subtyping methods often rely on random sampling to construct negative pairs, which may inadvertently engender false negatives. To overcome these challenges, we propose a novel end-to-end self-supervised learning model named Decoupled Contrastive Multi-view Clustering with adaptive false negative elimination (DCMC). Specifically, DCMC adopts a multi-view clustering architecture that facilitates intra- and inter-view contrastive learning across distinct embedding spaces, allowing view-specific information to be preserved while maintaining cross-view consistency. We further introduce an adaptive false negative elimination framework to progressively screen potential false negatives. Finally, pseudo-label rectification is applied to enhance the quality of the learned representations and further refine the clustering process. DCMC is evaluated on 10 commonly used cancer datasets against 19 state-of-the-art methods, with experimental results validating its superior performance. In the Liver Hepatocellular Carcinoma case study, differential expression analysis is performed to identify potential biomarkers, while the cancer subtypes identified by DCMC are validated for their responses to specific therapeutic drugs. The datasets and source code for DCMC are available online at https://github.com/LinMengX/DCMC.Author summary: Cancer is a heterogeneous disease characterized by diverse etiologies and clinical features. Different cancer subtypes exhibit substantial variations in prognostic responses and treatment outcomes. Therefore, accurate subtype identification is crucial for effective diagnosis and personalized therapeutic interventions. However, existing methods for cancer subtyping struggle to capture the inter- and intra-view relationships in multi-omics data and often overlook the issue of false negatives in contrastive learning. To address these issues, we introduce DCMC, an end-to-end self-supervised learning model that employs a decoupled multi-view clustering architecture to preserve view-specific information while ensuring consistency across different omics. DCMC further enhances clustering quality by integrating an adaptive false negative elimination framework to progressively filter out misleading negative pairs, along with pseudo-label rectification to refine the learned representations. Through tests on 10 widely adopted cancer datasets, the experimental results consistently demonstrate that DCMC outperforms 19 state-of-the-art methods in terms of −log10 P-values and the number of enriched clinical labels. A case study on Liver Hepatocellular Carcinoma confirms the clinical relevance and therapeutic drug sensitivity of DCMC-identified subtypes.