Massive-Scale Gene Co-Expression Network Construction and Robustness Testing Using Random Matrix Theory

Scott M. Gibson, Stephen Ficklin, Sven Isaacson, Feng Luo, F. Alex Feltus, Melissa C. SmithPLoS 2013 · DOI

Introduction

In systems biology, gene co-expression networks are vital for uncovering how genes interact to drive biological functions. By identifying genes that are frequently active simultaneously, researchers can infer shared roles in biological processes. However, two major challenges have historically hindered large-scale network analysis: the "thresholding problem"—distinguishing biological signal from random noise—and the "scalability problem"—the immense computational cost of mathematically rigorous thresholding methods like Random Matrix Theory (RMT).

The RMTGeneNet Solution

To address these hurdles, researchers developed RMTGeneNet, a high-performance C-based implementation of the RMT algorithm. RMT identifies the natural transition point where the distribution of "spacings" between eigenvalues in a correlation matrix shifts from random noise (Poisson) to ordered signal (Gaussian). While mathematically sound, previous implementations were too slow for robustness testing, which requires building hundreds of networks. RMTGeneNet offers a nearly 30x speedup, reducing construction time for large networks from days to hours.

Experimental Design and Datasets

The study evaluated the functional robustness of networks across three diverse organisms: humans, rice, and budding yeast. Using massive microarray datasets from the NCBI Gene Expression Omnibus (GEO), the team constructed "global" networks and then tested their resilience by creating "perturbed" versions—randomly removing 25%, 50%, or 75% of the input samples or probe sets.

Table 1: Summary of microarray datasets and genome assemblies used for network construction across human, rice, and yeast.

Functional Robustness vs. Topological Sensitivity

A key finding of the study is that biological function is more resilient than network topology. While removing samples might change specific "edges" (individual gene-to-gene connections), the underlying functional modules—grouped using the Link-Community Method (LCM)—remain remarkably stable.

The study also highlighted a "hub vulnerability." Networks are significantly more sensitive to the removal of genes (probe sets) than the removal of samples. Deleting highly connected "hub" genes can cause structural collapse, whereas reducing the number of samples used to calculate those connections has a much smaller impact on the overall functional map.

Figure A4: Number of shared nodes across networks as the percentage of samples increases, using 100% of available probe sets.

Key Results

  • Performance Leap: RMTGeneNet achieved a 29x speedup, cutting processing time for a typical network from 58 hours to just 2 hours.
  • Data Efficiency: Optimized storage formats reduced the data footprint by up to 90% (e.g., shrinking a rice dataset from 34GB to 5GB).
  • High Functional Similarity: Even with only 25% of original samples, functional similarity (measured by Kappa scores) remained high, often exceeding 0.6 for human datasets.
  • Diminishing Returns: For human networks, 300–500 samples were found to be a "sweet spot," capturing the majority of robust relationships without requiring thousands of additional data points.

Conclusion and Future Directions

RMTGeneNet enables the rapid construction and validation of large-scale gene networks. While the current tool is a significant advancement, future work will explore GPGPU acceleration to handle datasets exceeding 100,000 probe sets. This research confirms that RMT-thresholded networks capture biological signals that are fundamentally robust to data perturbations, providing a reliable foundation for systems biology research.

Summary updated

Abstract

The study of gene relationships and their effect on biological function and phenotype is a focal point in systems biology. Gene co-expression networks built using microarray expression profiles are one technique for discovering and interpreting gene relationships. A knowledge-independent thresholding technique, such as Random Matrix Theory (RMT), is useful for identifying meaningful relationships. Highly connected genes in the thresholded network are then grouped into modules that provide insight into their collective functionality. While it has been shown that co-expression networks are biologically relevant, it has not been determined to what extent any given network is functionally robust given perturbations in the input sample set. For such a test, hundreds of networks are needed and hence a tool to rapidly construct these networks. To examine functional robustness of networks with varying input, we enhanced an existing RMT implementation for improved scalability and tested functional robustness of human (Homo sapiens), rice (Oryza sativa) and budding yeast (Saccharomyces cerevisiae). We demonstrate dramatic decrease in network construction time and computational requirements and show that despite some variation in global properties between networks, functional similarity remains high. Moreover, the biological function captured by co-expression networks thresholded by RMT is highly robust.
BibTeX
@inproceedings{gibson2013massive,
  author = {Scott M. Gibson and Stephen Ficklin and Sven Isaacson and Feng Luo and F. Alex Feltus and Melissa C. Smith},
  title = {{Massive-Scale Gene Co-Expression Network Construction and Robustness Testing Using Random Matrix Theory}},
  booktitle = {PLoS},
  year = {2013},
  doi = {10.1371/journal.pone.0055871},
}
Figure