metasnf - Meta Clustering with Similarity Network Fusion
Framework to facilitate patient subtyping with similarity
network fusion and meta clustering. The similarity network
fusion (SNF) algorithm was introduced by Wang et al. (2014) in
<doi:10.1038/nmeth.2810>. SNF is a data integration approach
that can transform high-dimensional and diverse data types into
a single similarity network suitable for clustering with
minimal loss of information from each initial data source. The
meta clustering approach was introduced by Caruana et al.
(2006) in <doi:10.1109/ICDM.2006.103>. Meta clustering involves
generating a wide range of cluster solutions by adjusting
clustering hyperparameters, then clustering the solutions
themselves into a manageable number of qualitatively similar
solutions, and finally characterizing representative solutions
to find ones that are best for the user's specific context.
This package provides a framework to easily transform
multi-modal data into a wide range of similarity network
fusion-derived cluster solutions as well as to visualize,
characterize, and validate those solutions. Core package
functionality includes easy customization of distance metrics,
clustering algorithms, and SNF hyperparameters to generate
diverse clustering solutions; calculation and plotting of
associations between features, between patients, and between
cluster solutions; and standard cluster validation approaches
including resampled measures of cluster stability, standard
metrics of cluster quality, and label propagation to evaluate
generalizability in unseen data. Associated vignettes guide the
user through using the package to identify patient subtypes
while adhering to best practices for unsupervised learning.