Abstract:
|
Many multi-omics integration methods for disease subtyping have been proposed, with an intuition that incorporating more types of omics data produces better results. However, there are situations when integrating more omics data may negatively impact the performance of integration methods, such as including data type that is noisy in disease subtyping and redundant data types that may dominate the disease subtyping signals and decrease the subtyping accuracy. We propose a novel multi-omics regularized spectral clustering framework to integrate different omics data types, which learns the weights of each data type’s signal for disease subtyping as well as the signal’s redundancy level. Simulation studies and applications to multi-omics data of several cancer types from The Cancer Genome Atlas project suggest that the proposed multi-omics regularized spectral clustering framework achieves higher clustering accuracies and identifies new cancer subtypes that more accurately predict patient survival and are more biologically meaningful.
|