Differential Networks - Usage Guide

Overview

Compare co-expression and regulatory networks between conditions to find rewired gene-gene relationships. DiffCorr tests correlation differences with Fisher's z (gained/lost/reversed edges); DiffCoEx finds rewired modules; DINGO/iDINGO test direct (partial-correlation) rewiring; CoDiNA compares many networks. The key insight: differential connectivity is NOT differential expression -- a gene can rewire its partners with no change in mean expression and be the key signal (Hudson's myostatin). The dominant pitfall is power: rewiring needs more samples than DE, and pairwise testing has a p^2/2 multiple-testing explosion, so most small-cohort "rewired hub" claims are underpowered noise.

Prerequisites

# R packages
install.packages('DiffCorr')
install.packages('igraph')

# DGCA (archived from CRAN 2024-05, GitHub-only)
devtools::install_github('andymckenzie/DGCA')

# Python alternative
pip install pandas numpy scipy statsmodels networkx matplotlib

Minimum 20 samples per condition recommended (15 absolute floor).

Quick Start

Tell your AI agent what you want to do: - "Compare co-expression networks between disease and control" - "Find genes with rewired regulatory connections between conditions" - "Which gene pairs gain or lose co-expression in my treatment group?" - "Identify hub genes with the most differential connections"

Example Prompts

DiffCorr Analysis

"I have normalized RNA-seq data from 25 disease and 25 control samples. Find differentially correlated gene pairs."

"Run DiffCorr to compare co-expression between treated and untreated groups."

Network Comparison

"Which genes have the most rewired connections between my tumor and normal samples?"

"Show me gained, lost, and reversed edges in the differential network."

Direct vs Module-Level Rewiring

"I want direct (not indirect) rewiring between conditions. Run DINGO on partial correlations."

"Find modules that rewire between conditions with DiffCoEx instead of testing every edge."

Visualization

"Visualize the differential co-expression network colored by edge type."

"Create a network plot showing rewired connections for the top 50 most differentially connected genes."

What the Agent Will Do

Separate expression data by condition
Filter to top variable genes (2000-5000)
Compute correlation matrices for each condition
Test differential correlation using Fisher's z-transform
Classify edges as gained, lost, reversed, or unchanged
Apply FDR correction and identify rewired hub genes
Visualize differential network

Tips

Connectivity is not expression - report differential connectivity and DE separately; a non-DE gene can be the key rewired hub (Hudson's myostatin).
Power is the binding constraint - rewiring needs more samples than DE; below ~15-20 per group, results are mostly noise. Treat small-cohort findings as exploratory.
Tame the p^2/2 explosion - pre-filter to variable genes and apply strict FDR, or use module-level DiffCoEx to avoid per-edge testing entirely.
Marginal vs direct - DiffCorr gained edges are marginal (may reflect a shifted common driver); use DINGO partial correlations when directness matters.
FDR method - statsmodels multipletests defaults to Holm-Sidak; pass method='fdr_bh' for BH.
Effect-size filter - require absolute correlation change > 0.3 on top of significance to avoid trivial differences.
DGCA archived - removed from CRAN May 2024; install from GitHub if needed.

coexpression-networks - build the per-condition networks being compared
grn-inference - VIPER differential protein activity between conditions
scenic-regulons - TF regulon activity differences as a complementary readout
differential-expression/de-results - differential expression of means (the orthogonal question)
pathway-analysis/go-enrichment - functional enrichment of rewired gene sets
temporal-genomics/temporal-grn - time-resolved network change across stages