TPSC: a module detection method based on topology potential and spectral clustering in weighted networks and its application in gene co-expression module discovery

Yusong Liu, Xiufen Ye, Christina Y. Yu, Wei Shao, Jie Hou, Weixing Feng, Jie Zhang, Kun HuangBMC Bioinformatics 2021 · DOI

Abstract

Identifying functional modules in gene co-expression networks is a cornerstone of bioinformatics, yet traditional methods like WGCNA and lmQCM often struggle with imbalanced cluster sizes or low data coverage. This paper introduces TPSC (Topology Potential-based Spectral Clustering), a novel algorithm that integrates physics-inspired topology potential with spectral clustering. TPSC ensures 100% gene coverage and provides superior granularity, allowing for the discovery of disease-relevant pathways that are often obscured by broader methods. Applied to breast cancer datasets, TPSC demonstrated a higher capacity for identifying survival-significant modules and integrating prior biological knowledge.

Introduction

In the complex landscape of bioinformatics, identifying "modules"—groups of genes that work together—is essential for understanding biological systems. While tools like Weighted Gene Co-expression Network Analysis (WGCNA) have been industry standards, they frequently encounter the "giant component" problem, where a single massive cluster dominates the results, obscuring fine-grained biological signals. Conversely, methods like lmQCM provide finer detail but often discard a significant portion of the input data as noise.

TPSC (Topology Potential-based Spectral Clustering) addresses these limitations by providing a balanced approach that ensures full data coverage while maintaining high granularity. Furthermore, it offers a framework to incorporate existing medical knowledge into the clustering process, a feature often missing in purely unsupervised algorithms.

Methodology

The TPSC algorithm merges two powerful concepts to map weighted networks:

  1. Topology Potential: Borrowed from field theory in physics, this treats genes as particles with "mass." Each node exerts an influence on its neighbors based on their correlation (distance). This creates a "potential field" where cluster centers emerge as local maxima.
  2. Spectral Clustering: TPSC utilizes a "topology potential energy-corrected" Laplacian matrix. This mathematical adjustment ensures the matrix remains symmetric even when genes are assigned different masses, preventing computational errors common in other weighted methods.
Workflow of the TPSC algorithm to detect co-expression modules.

The algorithm is implemented in two versions: TPSC-1, which considers only connection strength, and TPSC-w, which allows researchers to assign weights to specific genes based on prior disease knowledge (e.g., Gene-Disease Association scores from databases like DisGeNET).

Parameter Selection and Granularity

A critical step in TPSC is determining the number of clusters ($k$). This is achieved through a greedy search for local maxima in the topology potential field. As shown in the relationship between the distance parameter ($d_{min}$) and $k$, the algorithm allows for a flexible determination of network granularity.

The relationship between the number of clusters (k) and the distance parameter (d_min) for TPSC-1 and TPSC-w.

Key Insights and Results

Testing on the TCGA-BRCA (The Cancer Genome Atlas Breast Cancer) dataset revealed several advantages of the TPSC method:

  • Balanced Module Sizes: Unlike WGCNA, which produced a giant module of 1,678 genes, TPSC generated more uniform and interpretable clusters. This prevents biological signals from being "swallowed" by a single massive component.
  • Superior Coverage: TPSC achieves 100% gene coverage. Every gene is assigned to a functional neighborhood, ensuring that potentially relevant data is not discarded.
  • Biological Precision: TPSC successfully separated broad biological processes into distinct sub-modules. For instance, it could distinguish between up-regulated and down-regulated components of the cell cycle, which WGCNA often lumps together.
  • Clinical Relevance: TPSC-1 identified four modules with significant differences in overall patient survival, outperforming both WGCNA and lmQCM, which identified only two each. Functional enrichment showed these modules were highly specific to critical pathways like P53 and PI3K-Akt signaling.

Conclusion

TPSC represents a significant advancement in gene co-expression network analysis. By combining physics-inspired topology potential with spectral clustering, it offers a more precise and inclusive lens for mapping biological networks. While the performance of the weighted version (TPSC-w) depends on the quality of prior knowledge, the overall framework provides a robust tool for precision medicine and general network analysis. Future applications may extend beyond bioinformatics into social and financial network modeling.

Summary updated

BibTeX
@article{liu2021tpsc,
  author = {Yusong Liu and Xiufen Ye and Christina Y. Yu and Wei Shao and Jie Hou and Weixing Feng and Jie Zhang and Kun Huang},
  title = {{TPSC: a module detection method based on topology potential and spectral clustering in weighted networks and its application in gene co-expression module discovery}},
  journal = {BMC Bioinformatics},
  year = {2021},
  doi = {10.1186/s12859-021-03964-5},
}
Figure