Frequency and Nonlinear Correlation Hybrid Network Inference
Fancy (Frequency And Nonlinear Correlation hYbrid) infers association networks by combining k-nearest neighbour mutual information (via MRNET) and distance correlation into a single hybrid edge score. Bootstrap resampling provides frequency-based confidence, and a weighted scoring function ranks edges for downstream module analysis. While designed for microbial co-abundance networks from metagenome-assembled genome (MAG) count tables, Fancy works with any numeric feature matrix, including gene co-expression, metabolite correlations, and cross-omics (e.g. gene-metabolite) associations.
# From GitHub
devtools::install_github("wala-github/Fancy")
# From Bioconductor (once accepted)
BiocManager::install("Fancy")Fancy/
├── R/ Core source code (14 files)
│ ├── fancy.R Main pipeline wrapper
│ ├── FancyResult-class.R S4 result class, validity, show/$ methods
│ ├── bootstrap.R Parallel bootstrap network inference (snowfall)
│ ├── mi_network.R Mutual information matrices (kNN estimator)
│ ├── dcor_network.R Distance correlation matrices (energy::dcor)
│ ├── scoring.R Edge frequency and hybrid score computation
│ ├── preprocess.R MAG filtering by coverage and prevalence
│ ├── parse_count_tables.R Sample name cleaning
│ ├── parse_taxonomy.R GTDB taxonomy parsing
│ ├── plot.R S4 plot method + network/elbow plotting
│ ├── export.R Cytoscape export with phyla colour palette
│ ├── utils.R Internal helpers
│ ├── data.R Dataset documentation
│ └── Fancy-package.R Package-level documentation
│
├── man/ roxygen2-generated documentation (.Rd files)
│
├── data/ Bundled example data (100 MAGs x 321 samples),
│ │ one dataset per .rda file:
│ ├── fancy_tiny_clr.rda
│ ├── fancy_tiny_counts.rda
│ ├── fancy_tiny_coverage.rda
│ ├── fancy_tiny_taxonomy.rda
│ └── fancy_tiny_metadata.rda
│
├── inst/extdata/ Reserved for external example data files
│
├── tests/ testthat test suite
│ ├── testthat.R Test harness
│ └── testthat/ Test scripts
│
├── vignettes/ Package vignettes
│ └── Fancy_workflow.Rmd HTML vignette (BiocStyle) walking through
│ the full hybrid network inference pipeline
│
├── DESCRIPTION Package metadata and dependencies
└── NAMESPACE Exports and imports (14 exported functions)
Standard correlation methods (e.g. Pearson, Spearman) miss nonlinear microbial interactions. Fancy's hybrid score captures patterns that linear metrics cannot, including threshold effects, context-dependent relationships, and feedback loops between MAGs.
| Threshold competitive exclusion | L-shaped nonlinear relationship | Context-dependent interaction |
|---|---|---|
| Bulleidia vs AC2028: Pearson r = -0.47 but the relationship is nonlinear with a clear threshold. Fancy hybrid score = 0.40. | RUG023 vs Cryptobacteroides: Pearson r = -0.31 underestimates a strong L-shaped dependency. Fancy hybrid score = 0.25. | CAG-791 vs Methanobrevibacter: Pearson r = -0.02 (no linear signal), yet Fancy hybrid score = 0.32 reveals a hidden association. |
Points are coloured by methane emission level (CH4 g/day: blue = low, yellow = mid, red = high), highlighting environment-dependent structure within these interactions.
library(Fancy)
# Run the full pipeline on a CLR-normalised count table
# (~15 minutes on the example data with 100 bootstraps and 4 CPUs;
# runtime scales with dataset size and number of bootstraps)
result <- fancy(t(fancy_tiny_clr), n_bootstrap = 100, cpus = 4)
# Inspect the result
# Show the hybrid score distribution
plot(result)
# Export for Cytoscape
data(fancy_tiny_taxonomy)
export_cytoscape(result, fancy_tiny_taxonomy, file_prefix = "my_network")GPL (>= 3)
If you use this software, please cite:
Fancy v0.99.0
DOI: https://doi.org/10.5281/zenodo.20003547