Principal Subspace Updation for Integrative Clustering of Multimodal Omics Data

Cancer subtyping is a key step towards the design of improved personalized therapies. Subtype discovery from large-scale multimodal data sets poses several challenges like data heterogeneity and high dimensionality. Moreover, existing integrative clustering algorithms tend to consider that each modality provides homogeneous and consistent subtype information, which may not be true for real life omics data sets. In this regard, this paper presents a fast algorithm to extract a low-rank joint subspace from the principal subspace of each individual modality such that the joint subspace best preserves the underlying subtype structure. The algorithm evaluates the quality of cluster information provided by each modality and the concordance of information shared among different modalities. This allows the algorithm to judiciously select the most relevant modalities and discard modalities providing noisy and inconsistent information while construction of the joint subspace. The performance of clustering in the joint subspace extracted by the proposed algorithm and its computational efficiency is compared with several existing integrative clustering approaches, on real life multimodal omics data sets. Moreover, survival analysis shows that the subtypes identified by the proposed approach have significantly different survival profiles.