moCluster: Identifying Joint Patterns Across Multiple Omics Data Sets

Increasingly, multiple omics approaches
are being applied to understand
the complexity of biological systems. Yet, computational approaches
that enable the efficient integration of such data are not well developed.
Here, we describe a novel algorithm, termed moCluster, which discovers
joint patterns among multiple omics data. The method first employs
a multiblock multivariate analysis to define a set of latent variables
representing joint patterns across input data sets, which is further
passed to an ordinary clustering algorithm in order to discover joint
clusters. Using simulated data, we show that moCluster’s performance
is not compromised by issues present in iCluster/iCluster+ (notably,
the nondeterministic solution) and that it operates 100× to 1000×
faster than iCluster/iCluster+. We used moCluster to cluster proteomic
and transcriptomic data from the NCI-60 cell line panel. The resulting
cluster model revealed different phenotypes across cellular subtypes,
such as doubling time and drug response. Applying moCluster to methylation,
mRNA, and protein data from a large study on colorectal cancer patients
identified four molecular subtypes, including one characterized by
microsatellite instability and high expression of genes/proteins involved
in immunity, such as PDL1, a target of multiple drugs currently in
development. The other three subtypes have not been discovered before
using single data sets, which clearly illustrates the molecular complexity
of oncogenesis and the need for holistic, multidata analysis strategies.