Hierarchical block matrix approach for multi-view clustering

Scientists are facing two important challenges when investigating life processes. First, biological systems, from gene regulation to physiological mechanisms, are inherently multiscale. Second, complex disease data collection is an expensive process, and yet the analyses are presented in a rather empirical and sometimes simplistic way, completely missing the opportunity of uncovering patterns of predictive relationships and meaningful profiles. In this work, we propose a multi-view clustering methodology that, although quite general, could be used to identify patient subgroups, for different omic information, by studying the hierarchical structures of the patient data in each view and merging their topologies. We first demonstrate the ability of our method to identify hierarchical structures in synthetic data sets and then apply it to real multi-view multi-omic data sets. Our results, although preliminary, suggest that this methodology outperforms single-view clustering approaches and could open several directions for improvements.