WIPER

WIPER is a Python implementation of Weighted In-Path Edge Ranking for ranking biomolecular associations in weighted networks.

This repository mirrors the layout of aimed-lab/WINNER: the installable Python package lives in python/, with root-level project notes and release history here.

The implementation follows:

• Yue Z, Nguyen T, Zhang E, Zhang J, Chen JY. WIPER: Weighted In-Path Edge Ranking for biomolecular association networks. Quantitative Biology. 2019;7(4):313-326. doi:10.1007/s40484-019-0180-y
• Local WIPER manuscript, presentation, poster, and historical result files used as implementation guidance.

Quick Start

pip install ./python

wiper --interactions interactions.tsv -o wiper_edges.tsv \
      --sigma 0.2 --iterations 200 --device auto --n-jobs -1 --include-novel

wiper-web

Input is a tab-delimited edge list. The first two columns are endpoint names; the third column is an edge weight. Weights are coerced to numeric values and clipped to [0, 1].

The output includes WIPER scores at initialization and after the requested iteration count, UFC/logUFC scores, p-values, significance symbols, ranks, and an extended flag for inferred novel edges.

Python API

from wiper import read_interactions, run_wiper1, run_wiper2

interactions = read_interactions("interactions.tsv")
result = run_wiper1(interactions, sigma=0.2, iterations=200, include_novel=True)
result.to_frame().to_csv("wiper_edges.tsv", sep="\t", index=False)

# Path-aware WIPER2 variant:
path_result = run_wiper2(interactions, sigma=0.85, iterations=200)

See python/README.md for implementation details, parallel CPU/GPU performance notes, and test commands.

Algorithm Variants

run_wiper1 preserves the published WIPER idea: compute an optimal-path node matrix D, transform endpoint distances into an edge-to-edge matrix X, then diffuse scores on X.

run_wiper2 is a path-aware variant: compute which shortest weighted paths actually traverse each edge, build a sparse edge-by-node-pair credit matrix WP, construct an edge co-path graph WP @ WP.T, and run WINNER-style restart propagation over that edge graph. This is often the more direct model if the scientific question is "which associations carry optimal network flow?"

WIPER2 is free for non-commercial research, education, evaluation, and academic use. Commercial use requires a separate written license granted by Dr. Jake Chen or another authorized copyright holder. See LICENSE-WIPER2-NONCOMMERCIAL.md.

The web explorer (wiper-web) runs a local browser-facing tool for loading or generating small networks, toggling raw/WIPER1/WIPER2 edge scores, inspecting rank tables, and trimming the plot to a top-N or top-percent backbone.

WIPER2 Performance

WIPER2 parallelizes the expensive shortest-path edge-credit construction over source-node blocks with joblib process workers (--n-jobs -1 uses all detected CPU cores). The final WINNER-style restart phase can dispatch to PyTorch CUDA or Apple MPS when available, although the current benchmark machine exposed only CPU.

On an Apple M4 with 10 detected CPU cores, a synthetic 300-node / 900-edge scale-free network improved from 1.33 s with one worker to 0.57 s with 10 workers for a full 200-iteration WIPER2 run. See python/README.md and python/benchmarks/benchmark_wiper2.py for the detailed table and reproducible benchmark command.