metasnf is a package that facilitates usage of the meta clustering paradigm described in (Caruana et al. 2006) with the similarity network fusion (SNF) data integration procedure developed in (Wang et al. 2014). The package offers a comprehensive suite of tools to assist users in transforming raw patient data into patient subtypes, decision making in the subtyping process, and visualization along the way with a strong emphasis on context-specific utility and principled validation of results.
You will need R version 4.1.0 or higher to install this package. We recommend installing the latest development version of the package (it has the best stuff!):
# Always get the latest version
devtools::install_github("BRANCHlab/metasnf")
# Install a specific tagged version
devtools::install_github("BRANCHlab/metasnf@v0.5.0")An extremely minimal usage of the package looks something like this:
# Load the package
library(metasnf)
# Setting up the data
data_list <- generate_data_list(
list(abcd_cort_t, "cortical_thickness", "neuroimaging", "continuous"),
list(abcd_cort_sa, "cortical_surface_area", "neuroimaging", "continuous"),
list(abcd_subc_v, "subcortical_volume", "neuroimaging", "continuous"),
list(abcd_income, "household_income", "demographics", "continuous"),
list(abcd_pubertal, "pubertal_status", "demographics", "continuous"),
uid = "patient"
)
# Specifying 5 different sets of settings for SNF
settings_matrix <- generate_settings_matrix(
data_list,
nrow = 5,
max_k = 40,
seed = 42
)
#> [1] "The global seed has been changed!"
# This matrix has clustering solutions for each of the 5 SNF runs!
solutions_matrix <- batch_snf(data_list, settings_matrix)
#> [1] "Row: 1/5 | Time remaining: 1 seconds"
#> [1] "Row: 2/5 | Time remaining: 1 seconds"
#> [1] "Row: 3/5 | Time remaining: 0 seconds"
#> [1] "Row: 4/5 | Time remaining: 0 seconds"
#> [1] "Row: 5/5 | Time remaining: 0 seconds"
#> [1] "Total time taken: 1 seconds."Check out the tutorial vignettes below to learn about how the package can be used:
And more tutorials can be found under the “articles” section of the documentation home page: https://branchlab.github.io/metasnf/index.html
Why use meta clustering?
Clustering algorithms seek solutions where members of the same cluster are very similar to each other and members of distinct clusters are very dissimilar to each other. In sufficiently noisy datasets where many qualitatively distinct solutions with similar scores of clustering quality exist, it is not necessarily the case that the top solution selected by a clustering algorithm will also be the most useful one for the user’s context.
To address this issue, the original meta clustering procedure Caruana et al., 2006 involved generating a large number of reasonable clustering solutions, clustering those solutions into qualitatively similar ones, and having the user examine those “meta clusters” to find something that seems like it’ll be the most useful.
Why use SNF?
In the clinical data setting, we often have access to patient data across a wide range of domains, such as imaging, genetics, biomarkers, demographics. When trying to extract subtypes out of all this information, direct concatenation of the data followed by cluster analysis can result in a substantial amount of lost (valuable) signal contained in each individual domain. Empirically, SNF has been demonstrated to effectively integrate highly diverse patient data for the purposes of clinical subtyping.
- Stability measures and consensus clustering
- Removing unwanted signal
- Parallel processing
- Label propagation
Caruana, Rich, Mohamed Elhawary, Nam Nguyen, and Casey Smith. 2006. “Meta Clustering.” In Sixth International Conference on Data Mining (ICDM’06), 107–18. https://doi.org/10.1109/ICDM.2006.103.
Wang, Bo, Aziz M. Mezlini, Feyyaz Demir, Marc Fiume, Zhuowen Tu, Michael Brudno, Benjamin Haibe-Kains, and Anna Goldenberg. 2014. “Similarity Network Fusion for Aggregating Data Types on a Genomic Scale.” Nature Methods 11 (3): 333–37. https://doi.org/10.1038/nmeth.2810.