metaSNF: Meta clustering with Similarity Network Fusion
Brief Overview
metaSNF is an R 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 multi-modal tabular data into cluster solutions, decision making in the clustering process, and visualization along the way with a strong emphasis on context-specific utility and principled validation of results.
Installation
You will need R version 4.1.0 or higher to install this package. metaSNF can be installed from CRAN:
install.packages("metasnf")Development versions can be installed from GitHub:
# Latest development version
devtools::install_github("BRANCHlab/metasnf")
# Install a specific tagged version
devtools::install_github("BRANCHlab/metasnf@v2.1.1")Quick Start
Minimal usage of the package looks like this:
# Load the package
library(metasnf)
# Setting up the data
dl <- 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"
)
#> ℹ 176 observations dropped due to incomplete data.
# Specifying 5 different sets of settings for SNF
set.seed(42)
sc <- snf_config(
dl,
n_solutions = 5,
max_k = 40
)
#> ℹ No distance functions specified. Using defaults.
#> ℹ No clustering functions specified. Using defaults.
# This matrix has clustering solutions for each of the 5 SNF runs!
sol_df <- batch_snf(dl, sc)
sol_df
#> 5 cluster solutions of 100 observations:
#> solution nclust mc uid_NDAR_INV0567T2Y9 uid_NDAR_INV0IZ157F8
#> 1 2 . 1 2
#> 2 2 . 2 1
#> 3 10 . 1 9
#> 4 2 . 2 1
#> 5 8 . 1 7
#> 98 observations not shown.
#> Use `print(n = ...)` to change the number of rows printed.
#> Use `t()` to view compact cluster solution format.
t(sol_df)
#> 5 cluster solutions of 100 observations:
#> uid s1 s2 s3 s4 s5
#> uid_NDAR_INV0567T2Y9 1 2 1 2 1
#> uid_NDAR_INV0IZ157F8 2 1 9 1 7
#> uid_NDAR_INV0J4PYA5F 1 2 7 2 3
#> uid_NDAR_INV10OMKVLE 2 1 8 1 4
#> uid_NDAR_INV15FPCW4O 1 2 2 2 5
#> uid_NDAR_INV19NB4RJK 1 2 9 2 7
#> uid_NDAR_INV1HLGR738 1 2 9 2 7
#> uid_NDAR_INV1KR0EZFU 2 1 9 2 7
#> uid_NDAR_INV1L3Y9EOP 1 2 2 2 5
#> uid_NDAR_INV1TCP5GNM 1 2 8 2 4
#> 90 observations not shown.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
Background
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.
Documentation
Example workflows
Essential objects
Further customization of generated solutions
Additional functionality
- Stability measures and consensus clustering
- Removing unwanted signal
- Parallel processing
- Label propagation
- Imputations
- NMI scores
Plotting
References
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.