Learning an Adaptive Self-expressive Fusion Model for Multi-omics Cancer Subtype Prediction

The discovery of cancer subtypes has helped researchers gain deeper insights into the study of oncology heterogeneity. However, since cancer complexity exists at various omics levels, extracting and fusing complementary information across multi-omics are still challenges. We proposed an Adaptive Self-Expressive Fusion Model (ASEFM) based on attention and consistency constraints to predict cancer subtypes. Firstly, the self-expression network is employed to compute the coefficient matrix for each omics. Secondly, ASEFM projects the most correlated similarity into a shared feature space and assigns weights to every specific similarity feature through an attention mechanism. Finally, the multi-omics self-expressive coefficients are calculated from common similarity features and specific similarity features under the constraints of consistency and disparity. Ten cancer datasets from the cancer genome atlas platform were used for model performance evaluation. The survival analysis and function enrichment analysis results indicate that the cancer subtype identified by ASEFM holds significant biological significance.