Integrative multi-omics cancer subtyping via adaptive neighborhood kernel learning and survival-evaluated subspace clustering

The integration of multi-omics data offers transformative potential for precision oncology by enabling comprehensive molecular characterization of tumors, yet existing clustering methods often struggle with the heterogeneous geometry, noise susceptibility, and complex manifold structures inherent to cross-omics data. We present Adaptive Neighborhood Multi-Kernel Subspace Clustering (ANMKSC), a framework that addresses these challenges through three complementary components: (1) swarm intelligence-guided kernel selection that dynamically optimizes cross-omics kernel interactions while mitigating technical artifacts; (2) neighborhood-focused kernel learning that preserves high-confidence local similarities to capture the intrinsic global manifold structure; and (3) survival-evaluated subspace segmentation whose outputs are independently validated against clinical outcomes for biologically coherent cancer subtyping. Across repeated stochastic runs, ANMKSC identifies computational subtypes with distinct retrospective molecular and overall-survival associations. Across the evaluated datasets, it achieves competitive performance and the best survival-stratified results in most of the reported clustering settings. These findings provide a basis for further validation in independent cohorts.