Perturbation efficiency resolves target-count bias in network proximity metrics: A controlled audit

This repository contains the complete, reproducible research pipeline, data, and manuscript source for our study identifying and resolving systematic bias in network medicine metrics. Through a controlled audit using the human liver interactome, we demonstrate that proximity-based Z-scores are confounded by target set size and provide perturbation efficiency as a stable, resolution-focused alternative.

---

Scientific Context

Network-based drug prioritization typically assumes that topological proximity reflects functional relevance. However, we demonstrate that the standard proximity Z-score is fundamentally confounded by the Law of Large Numbers (LLN): as target set size increases, the null distribution variance decreases, leading to deterministic significance inflation.

Using Hypericum perforatum (St. John's Wort) as a controlled model system, we pair a known biological ground truth (Hyperforin-mediated hepatotoxicity) with extreme target-count asymmetry (10 vs 62 targets). This study provides a methodological audit of this bias and proof-of-concept for its resolution via perturbation efficiency—a size-normalized influence metric.

Key Results (STRING >=900 Liver LCC)

Compound	Targets	Proximity ($d_c$)	Proximity Z	Influence Z (RWR)	Efficiency (Avg Inf)
Hyperforin	10	1.30	-3.86	+10.12	0.1138
Quercetin	62	1.68	-5.44	+4.55	0.0322

Important

Hyperforin achieves ~3.7x more directed influence per-target than Quercetin, correctly identifying the high-leverage modulator where proximity metrics fail. This stability is maintained across varying network thresholds (≥700 and ≥900) and expression-weighted environments.

---

Quick Start (Reproducibility)

1. Environment Setup

# Clone the repository
git clone https://github.com/antonybevan/h-perforatum-network-tox
cd h-perforatum-network-tox

# Install dependencies (Python & R required)
pip install -r requirements.txt

2. Run Analysis Pipeline

Execute the end-to-end Python analysis (Network construction, Permutations, Bootstrap):

python scripts/run_pipeline.py

3. Generate Publication Figures

Generate the figures for the manuscript using R:

source("R/fig2_dumbbell.R")
source("R/fig3_ewi_waterfall.R")

---

Repository Structure

├── src/network_tox/     # Core analytical modules (RWR, EWI, Permutation)
├── scripts/             # Production execution scripts
├── R/                  # Publication-tier plotting scripts
├── data/               # Curated target and DILI gene sets (DILIrank, DisGeNET)
├── results/            # Computed Z-scores and consolidated tables
├── manuscript/         # LaTeX source (Scientific Reports format) and final PDFs
├── tests/              # Validation suite for core algorithms

---

Methodology Summary

1. Network Construction: STRING v12.0 PPI (High confidence ≥900), filtered for liver-expressed genes (GTEx v8, TPM ≥1).
2. Permutation Testing: 1,000 degree-matched permutations per compound to control for topology-specific degree bias.
3. Perturbation Efficiency: Normalization of steady-state influence mass to target set size, enabling direct comparison across asymmetric polypharmacology.
4. Expression Weighting (EWI): Destination-node transition weighting based on tissue-specific protein abundance (GTEx liver TPM).
5. Bootstrap Sensitivity: Empirical validation via random subset sampling to exclude target-count artifact.

---

Citation

If you use this framework or the data sets, please cite:

@article{bevan2026perturbation,
  title={Perturbation efficiency resolves target-count bias in network proximity metrics: A controlled audit},
  author={Bevan, Antony},
  year={2026},
  journal={(under review)},
  url={https://github.com/antonybevan/h-perforatum-network-tox}
}

---

License

Distributed under the MIT License. See LICENSE for more information.