Part Mutual Information for Direct Association Network Inference • PartMI

{Part Mutual Information for Direct Association Networks}

Overview

PartMI implements Part Mutual Information (PMI) for inferring direct associations in biological networks. Unlike standard correlation or mutual information (MI), PMI conditions on background variables to remove indirect dependencies, substantially reducing false-positive edges.

Why PMI?

Method	Detects non-linear relationships	Removes indirect edges	Computationally feasible
Pearson correlation	No	No	Yes
Mutual Information (MI)	Yes	No	Yes
Partial correlation	No	Yes	Yes
Part Mutual Information (PMI)	Yes	Yes	Yes

Installation

# install.packages("remotes")
remotes::install_github("danymukesha/PartMI")

Quick Start

library(PartMI)

# Simulate expression data
set.seed(42)
n <- 200
z <- rnorm(n)                    # hidden regulator
x <- 0.7 * z + rnorm(n, 0, 0.5) # gene X
y <- 0.7 * z + rnorm(n, 0, 0.5) # gene Y (indirect via Z)

# MI picks up the indirect association
mi(x, y)  # > 0 (false positive)

# PMI removes the common cause Z
pmi(x, y, z)  # ~ 0 (correct)

Core Functions

Function	Description
`mi()`	Mutual information (binning or kNN estimator)
`pmi()`	Part/conditional mutual information X vs Y given Z
`pmi_network()`	Build full network with significance testing
`normalize_data()`	Z-score, rank, minmax, or quantile normalization
`discretize_data()`	Quantile or equal-width discretization

References

Zhao et al. (2016) Part mutual information for quantifying direct associations in networks. PNAS 113(18):5130-5135.
Frenzel & Pompe (2007) Partial Mutual Information for Coupling Analysis of Multivariate Time Series. Phys. Rev. Lett. 99:204101.
Kraskov, Stögbauer & Grassberger (2004) Estimating mutual information. Phys. Rev. E 69:066138.