Network Visualization - Usage Guide

Overview

Network visualizations render protein-interaction, gene-regulatory, co-expression, and pathway graphs. Layout choice dominates the result: force-directed (Fruchterman-Reingold, ForceAtlas2) is general-purpose but layout positions are NOT biologically meaningful; hive plots (Krzywinski 2012) and circular layouts preserve categorical structure; hierarchical edge bundling (Holten 2006) reduces hairball clutter for many-to-many. NetworkX + matplotlib for static publication; PyVis for interactive HTML; Cytoscape for journal-grade composition.

Prerequisites

pip install networkx pyvis py4cytoscape matplotlib numpy
# Large networks:
pip install datashader

install.packages(c('igraph', 'ggraph', 'tidygraph'))

Quick Start

Tell your AI agent what you want to do: - "Plot PPI network with degree-sized nodes and community-colored modules using NetworkX" - "ForceAtlas2 layout for a 5000-node scale-free network" - "Interactive PyVis HTML with hover tooltips" - "Send the network to Cytoscape with degree-based styling via py4cytoscape" - "Hive plot anchored by module assignment"

Example Prompts

Standard PPI

"PPI network of 200 proteins from STRING. Spring layout with fixed seed. Color nodes by community via greedy_modularity_communities. Size by degree. Label only hubs (degree >= 5)."

Large network

"5000-node co-expression network. Use ForceAtlas2 layout. Bundle edges with hierarchical edge bundling. Rasterize for PDF."

Comparing conditions

"Two PPI networks (control vs treatment). Compute layout on the UNION graph; render both with the same pos so visual changes reflect biology not layout."

Interactive supplement

"PyVis HTML with degree-based node size and module-based color; force-atlas physics; export standalone HTML."

Cytoscape pipeline

"Send NetworkX graph to Cytoscape via py4cytoscape; apply degree-mapping style; export PDF."

What the Agent Will Do

Load network from edge list / GraphML / SIF / interaction-database query.
Compute basic stats (degree distribution, density, modularity).
Choose layout: force-directed for general; hive for comparative; circular for small dense; edge bundling for many-to-many.
Set random seed (always); reuse pos when comparing conditions.
Compute node attributes (degree, betweenness, community assignment).
Render with NetworkX layered drawing (draw_networkx_edges, nodes, labels) for fine control.
Label hubs only (cap at degree threshold or ~30 labels).
Encode edge weights via line width when meaningful.
Export PDF for static (with raster for >2000 nodes) or HTML for interactive.

Tips

Layout positions are NOT biology. Frame conclusions on edge existence and node degree, not inter-node distance.
Always set seed. seed=42 in NetworkX, set.seed(42) for R igraph. Without it, layout varies across runs.
Comparing conditions requires shared layout. Compute on union network; reuse pos.
ForceAtlas2 for scale-free (>500 nodes, hub-spoke topology). Spring for general; Kamada-Kawai for small dense.
Hive plots (Krzywinski 2012) anchor nodes to axes by categorical attribute - biology-faithful alternative to force-directed.
Hierarchical edge bundling (Holten 2006) dramatically reduces clutter for many-to-many networks within a hierarchy.
Cap labels at hubs only (degree > 5 or genes of interest). Labeling all nodes destroys readability.
Encode edge weight via line width when interaction confidence varies - uniform widths hide information.
PyVis max ~2000 nodes for browser HTML. Above this use Cytoscape.js or Datashader.
Datashader for >50000 nodes - raster aggregation; only honest display at that scale.
Cytoscape (desktop) for publication-grade biological networks. py4cytoscape exposes script control from Python/R.

gene-regulatory-networks/coexpression-networks - Build the network
database-access/interaction-databases - Fetch PPI data
data-visualization/multipanel-figures - Combine with other plots
data-visualization/color-palettes - Community palette
single-cell/cell-communication - Cell-cell interaction networks