UnPaSt

UnPaSt is a novel method for identification of differentially expressed biclusters.

Requirements:

Python (version 3.8.16):
    fisher==0.1.9
    jenkspy==0.2.0
    pandas==1.3.5
    python-louvain==0.15
    matplotlib==3.7.1
    seaborn==0.11.1
    numba==0.51.2
    numpy==1.22.3
    scikit-learn==1.2.2
    scikit-network==0.24.0
    scipy==1.7.1
    statsmodels==0.13.2
    lifelines==0.27.4

R (version 4.3.1):
    WGCNA==1.70-3
    limma==3.42.2

Installation tips

It is recommended to use "BiocManager" for the installation of WGCNA:

install.packages("BiocManager")
library(BiocManager)
BiocManager::install("WGCNA")

Examples

• UnPaSt requires a tab-separated file with features (e.g. genes) in rows, and samples in columns. Feature and sample names must be unique.

cd test;
mkdir -p results;

# running UnPaSt with default parameters and example data
python ../run_unpast.py --exprs scenario_B500.exprs.tsv.gz --basename results/scenario_B500

# with different binarization and clustering methods
python ../run_unpast.py --exprs scenario_B500.exprs.tsv.gz --basename results/scenario_B500 --binarization ward --clustering Louvain

# help
python run_unpast.py -h

Outputs

• <basename>.[parameters].biclusters.tsv - a .tsv table with found biclsuters, where
  • the first line starts from '#' and stores parameters
  • each following line represents a bicluster
  • SNR column contains SNR of a bicluster
  • columns "n_genes" and "n_samples" provide the numbers of genes and samples, respectively
  • "gene","sample" contain gene and sample names respectively
  • "gene_indexes" and "sample_indexes" - 0-based gene and sample indexes in the input matrix.
• binarized expressions, background distributions of SNR for each bicluster size and binarization statistics [if clustering is WGCNA, or '--save_binary' flag is added]

About

UnPaSt is an unconstrained version of DESMOND method (repository, publication)

Major modifications:

• it does not require the network of feature interactions
• UnPaSt clusters individual features instead of pairs of features
• uses 2-means, hierarchicla clustering or GMM for binarization of individual gene expressions
• SNR threshold for featuer selection is authomatically determined; it depends on bicluster size in samples and user-defined p-value cutoff