Metabolite Protein Network Construction

metabolite-protein-network-construction

Summary

Construct and visualize bipartite network graphs linking metabolites to proteins, weighted by association strength and disease class, to reveal disease-specific metabolic–proteomic relationships from LC-MS data. This skill transforms raw correlation and prediction outputs into interpretable network topology suitable for publication.

When to use

Apply this skill after generating metabolite-disease correlation data and protein association predictions from a deep learning metabolomics module (e.g., DeepMSProfiler's feature extraction step). Use it when you need to communicate multi-disease metabolite–protein associations as a unified graph rather than as separate heatmaps or lists, or when the research question requires visual exploration of network topology, hub metabolites, or disease-specific connectivity patterns.

When NOT to use

• Input is already a publication-ready network image or has been manually curated for display — construction is unnecessary.
• No protein association predictions are available; network would be unipartite metabolite-only, which may be better served by alternative metabolite correlation network workflows.
• The research question focuses on statistical testing or quantitative comparison of correlations rather than visual network topology discovery.

Inputs

• metabolite-disease correlation matrix (numpy array or table)
• protein association prediction scores (array or dataframe)
• disease class labels (categorical array matching samples or metabolite groups)
• association strength metadata (edge weights, confidence scores)

Outputs

• bipartite network graph object (networkx or igraph format)
• network visualization image (PNG, SVG, or PDF file)
• node and edge coordinate data (optional, for reproducibility)

How to apply

Load the metabolite-disease correlation matrix and protein association predictions generated by the deep learning module. Construct a bipartite graph where metabolites and proteins form two distinct node sets, with edges weighted by correlation strength or prediction confidence and colored/stratified by disease class. Apply a force-directed or hierarchical layout algorithm (e.g., Fruchterman–Reingold) to position nodes for interpretability and minimize edge crossing. Render the network with disease-type color coding on nodes, node size scaled proportionally to association strength, and edge transparency or width reflecting confidence scores. Export the finalized network as a high-resolution vector or raster image (PNG, SVG, PDF) suitable for publication.

Related tools

• DeepMSProfiler (generates metabolite-disease correlations and protein association predictions as upstream inputs to network construction) — https://github.com/yjdeng9/DeepMSProfiler
• Python (networkx, igraph, matplotlib/seaborn) (libraries for graph construction, layout algorithms, and visualization rendering)

Examples

# After DeepMSProfiler feature extraction:
from DeepMSProfiler import run_feature, show_feature
run_feature(job_dir='../jobs/jobs007')
show_feature(job_dir='../jobs/jobs007', mode='ensemble')

Evaluation signals

• Verify bipartite structure: metabolite and protein nodes are in separate, non-overlapping sets; edges only cross between sets, never within.
• Check node scaling: node size correlation with association strength is monotonic and visually discriminable across at least 3–5 magnitude tiers.
• Confirm disease stratification: nodes or edges exhibit clear, consistent color or visual clustering by disease class; disease-specific subnetworks should be visually separable.
• Validate layout stability: run the layout algorithm 2–3 times with different random seeds; node positions should be qualitatively similar (same neighbor relationships preserved), indicating deterministic convergence.
• Confirm export quality: high-resolution output (≥300 DPI if raster, or vector format) is readable at publication size; edge labels and node IDs are not occluded.

Limitations

• Network layout algorithms (force-directed, hierarchical) are stochastic and sensitive to initialization; reproducibility requires fixed random seeds.
• Large networks (>1000 nodes) become visually cluttered; subgraph extraction or interactive visualization may be required for interpretability.
• Edge transparency and color encoding are limited to ~5–7 visually distinct categories; disease classes >7 or continuous confidence scores require binning or alternative encoding (e.g., edge width).
• DeepMSProfiler's protein association predictions are computed post-hoc from metabolite features; direct protein measurement data is not incorporated, potentially inflating apparent protein-metabolite relationships.

Evidence

• [other] Construct a bipartite network graph with metabolites and proteins as nodes, weighted by association strength and disease class.: "Construct a bipartite network graph with metabolites and proteins as nodes, weighted by association strength and disease class."
• [other] Apply network layout algorithm (force-directed or hierarchical) to position nodes for interpretability.: "Apply network layout algorithm (force-directed or hierarchical) to position nodes for interpretability."
• [other] Render the network plot with disease-type color coding, node size scaled by association strength, and edge transparency reflecting confidence.: "Render the network plot with disease-type color coding, node size scaled by association strength, and edge transparency reflecting confidence."
• [readme] Disease-associated metabolite-protein network plots: "Disease-associated metabolite-protein network plots"
• [other] Export the finalized network plot as a high-resolution image file suitable for publication.: "Export the finalized network plot as a high-resolution image file suitable for publication."