SeOMLR: one-step multi-view latent representation with self-weighted ensemble learning for multi-omics cancer...

Accurate cancer subtyping is critically important for cancer treatment due to significant molecular heterogeneity. While existing methods with multi-omics integration have achieved some success in cancer subtype identification by leveraging the rich information provided by multi-omics data, most approaches remain limited by an overemphasis on cross-omics consistency at the expense of intra-omics specificity. Furthermore, a two-step scheme is often adopted to extract cluster structure from a consistency matrix or a continuous indicator matrix by k-means, which inevitably leads to information loss and unstable clusters.

To overcome these issues, we propose seOMLR, a one-step multi-view latent representation method with self-weighted ensemble learning for cancer subtyping. Using relaxed exclusivity constraints and consistency regularization terms, seOMLR exploits the specificity and consistency of multi-omics data by building a sparse low-rank self-representation framework. Simultaneously, a self-weighted ensemble strategy is introduced to adaptively incorporate prior subtyping information from other methods, indirectly promoting specificity and consistency learning. Moreover, the discrete clustering structure is subsequently extracted via spectral rotation to avoid information loss and cluster instability. Through joint iterative optimization of fusion and clustering, seOMLR enhances subtyping accuracy. Experiments on both simulated datasets and eight real multi-omics cancer datasets from TCGA demonstrate that seOMLR outperforms competing methods, achieving efficient multi-omics data fusion and providing computational framework support for cancer subtyping research.

Supplementary data are available at Bioinformatics online.