SIGNET

Signed Network Community Detection (SIGNET) detects community structure in signed networks — networks with both positive (attractive) and negative (repulsive) edges. It extends classical community detection algorithms (Louvain, Leiden) to leverage negative edge information, which is critical for accurate module detection in biological networks such as gene co-expression networks.

Why signed networks?

Standard community detection methods use only positive edges and ignore negative correlations between genes. When inter-community positive edges are abundant (e.g., in correlation-based networks), unsigned methods fail to identify the correct community structure because they cannot distinguish "noise" positive edges from true intra-community edges.

Signed methods use negative edges as repulsive forces to push nodes apart, recovering the correct community structure even when the positive graph alone is ambiguous.

Signed adjacency matrix sorted by detected communities. As inter-community positive edges increase (top to bottom), unsigned modularity collapses (ARI=0.07) while signed methods maintain near-perfect recovery (ARI>0.93).

Methods

SIGNET provides three community detection methods for signed networks:

Method	Algorithm	Objective	Best for
louvain	LouvainSigned	α·Q⁺ − (1−α)·Q⁻	Compatibility with existing Louvain workflows
leiden-mod-alpha	LeidenSigned (modularity)	α·Q⁺ − (1−α)·Q⁻	General signed networks
leiden-cpm-single	LeidenSigned (CPM)	Σ(w_ij − γ) on signed graph	Correlation-based networks (co-expression)

When to use CPM over modularity: Modularity's null model assumes a configuration-model random graph. In correlation-based networks (e.g., gene co-expression), the positive graph is inherently dense, distorting the null model. CPM uses an absolute density threshold (γ) instead and is not affected by this issue.

Installation

pip install signet

pip install -i https://test.pypi.org/simple/ signet-gene==0.0.8

Dependencies

• Python ≥ 3.9
• numpy
• igraph (python-igraph)
• leidenalg
• networkx
• matplotlib (for plotting outputs)
• scipy (for MAT format support)

Quick start

# Basic run with Leiden signed modularity
signet --pos positive_edges.tsv --neg negative_edges.tsv \
  --method leiden-mod-alpha --alpha 0.6 --resolution 1.0 \
  --out-prefix results/my_analysis

# Leiden signed CPM (recommended for co-expression networks)
signet --pos positive_edges.tsv --neg negative_edges.tsv \
  --method leiden-cpm-single --gamma 0.05 --lambda-neg 1.0 \
  --out-prefix results/my_analysis

# Louvain signed
signet --pos positive_edges.tsv --neg negative_edges.tsv \
  --method louvain --alpha 0.6 --resolution 1.0 \
  --out-prefix results/my_analysis

Input formats

TSV (default)

Tab-separated file with five columns (no header):

gene_id1    gene_id2    gene_name1    gene_name2    weight
ENSG00001   ENSG00002   GeneA         GeneB         15.3
ENSG00001   ENSG00003   GeneA         GeneC         12.1

Provide separate files for positive and negative edges via --pos and --neg. Use --thre-pos and --thre-neg to filter edges by weight.

MAT (MATLAB sparse matrix)

signet --format mat --mat data/network.mat \
  --pos-key pos --neg-key neg \
  --method leiden-mod-alpha --out-prefix results/mat_run

The MAT file should contain two sparse matrices (positive and negative adjacency).

Output files

When --out-prefix is specified, SIGNET generates the following outputs:

File	Description
<prefix>_partition.tsv	Community assignment for each gene (gene_id, gene_name, community)
<prefix>_summary.txt	Parameters, graph statistics, community size distribution, entropy
<prefix>_communities.gmt	GMT format for direct use with GSEA, Enrichr, clusterProfiler
<prefix>_community_sizes.pdf	Rank-size plot and histogram of community sizes
<prefix>_inter_community.pdf	Heatmap of inter-community edge density (positive / negative / signed)
<prefix>_module_<id>.pdf	Subnetwork visualization for top modules (positive=red, negative=blue)

Disabling outputs

--no-plot     # Skip all PDF plots
--no-gmt      # Skip GMT output
--quiet       # Suppress progress messages

Controlling module plots

--top-modules 10             # Number of top modules to plot (default: 10)
--max-nodes-per-module 200   # Max nodes per module plot (default: 200)

Parameters

Method-specific parameters

Modularity-based methods (louvain, leiden-mod-alpha):

Parameter	Description	Default
--alpha	Balance between positive and negative modularity (0–1). Higher α emphasizes positive edges.	0.5
--resolution	Resolution parameter. Higher values produce more, smaller communities.	1.0

CPM method (leiden-cpm-single):

Parameter	Description	Default
--gamma	CPM resolution. Minimum edge density within communities.	0.5
--lambda-neg	Weight multiplier for negative edges. Controls how strongly negative edges repel.	0.0
--neg-weight-mode	How to transform negative weights: absolute (−λ·|w|) or signed (λ·w).	absolute

Common parameters

Parameter	Description	Default
--seed	Random seed for reproducibility.	None
--out-prefix	Output file prefix. If omitted, only prints summary to stdout.	None

Python API

SIGNET can also be used as a Python library:

import signet.network_module as nr
import signet.LeidenSigned as les

# Load graphs
G_pos = nr.load_graph_from_TSV_igraph("positive.tsv", threshold=10)
G_neg = nr.load_graph_from_TSV_igraph("negative.tsv", threshold=5)

# Leiden signed modularity
partition = les.find_partition_signed_modularity_alpha(
    G_pos, G_neg, alpha=0.6, resolution=1.0, seed=42
)

# Leiden signed CPM (single signed graph)
G_signed = nr.load_signed_graph_from_two_TSV_igraph(
    "positive.tsv", "negative.tsv",
    pos_threshold=10, neg_threshold=5, lambda_neg=1.0
)
partition = les.find_partition_signed_CPM_single_graph(
    G_signed, gamma=0.05, seed=42
)

# Inspect results
print(partition.membership)
nr.display_communities_by_name(G_pos, partition)
nr.count_nodes_in_communities(partition)

Visualization

# Visualize a specific module (positive=red, negative=blue)
nr.visualize_module_signed(G_pos, G_neg, partition, community_id=0)

# Visualize the module containing a specific gene
nr.visualize_module_of_gene_signed(G_pos, G_neg, partition, "TP53")

# Top-degree nodes within a module
nr.visualize_module_of_gene_top_degree_nodes_signed(
    G_pos, G_neg, partition, "TP53", top_n=30
)

Benchmarking

SIGNET includes a simulation framework for evaluating community detection methods on signed networks.

# Quick test (no signet package required)
python sim3_benchmark.py test

# Full benchmark with all methods and parameter grids
python sim3_benchmark.py run sim3_results/

# Generate example correlation matrix visualizations
python sim3_benchmark.py plot sim3_results/

Simulation types

Type	Model	Tests
Type A	Correlation-based co-expression	Modularity null model distortion on dense positive graphs
Type B	Planted partition (stochastic block model)	Control — balanced positive/negative structure
Type C	scRNA-seq exclusive expression	Multipartite negative graph where "enemy of enemy ≠ friend"

Example visualizations

# Generate concrete examples showing signed vs unsigned differences
python sim3_examples.py sim3_examples/

This produces network graphs and adjacency matrix heatmaps at three difficulty levels, clearly demonstrating when signed methods outperform unsigned methods.

How it works

Signed modularity (multiplex optimization)

The positive and negative graphs are treated as two layers of a multiplex network. The objective function is:

Q_signed = α · Q_modularity(G⁺) − (1 − α) · Q_modularity(G⁻)

This is optimized using leidenalg.optimise_partition_multiplex, which simultaneously considers both layers with different weights.

Signed CPM (single graph)

Positive and negative edges are combined into a single graph with signed weights:

w_signed(i,j) = w_pos(i,j) − λ_neg · w_neg(i,j)

Standard CPM is then applied, where the objective function naturally penalizes negative edges within the same community:

H_CPM = Σ_{i,j in same community} (w_signed(i,j) − γ)

Choosing a method

Is your network correlation-based (co-expression, etc.)?
├── Yes → Use leiden-cpm-single
│         Start with --gamma 0.05 --lambda-neg 1.0
│         Increase gamma for smaller communities
│
└── No (e.g., social network, citation network)
    ├── Need reproducibility/stability? → Use leiden-mod-alpha
    │   Start with --alpha 0.6 --resolution 1.0
    │
    └── Compatibility with existing Louvain pipeline? → Use louvain
        Start with --alpha 0.6 --resolution 1.0

Citation

If you use SIGNET in your research, please cite:

[Citation information to be added]