Comparison of single gene and module-based methods for modeling gene regulatory networks

Mikel Hernáez, Olivier GevaertbioRxiv 2018 · DOI

TL;DR

This paper introduces a module-based approach for inferring gene regulatory networks, which significantly outperforms traditional single-gene methods. By first identifying groups of co-regulated genes (modules) and then linking these modules to regulatory genes using a variational Bayes method, the authors achieve more informative and biologically relevant network models.

What Problem Does This Paper Address?

Reverse engineering gene regulatory networks (GRNs) from gene expression data is a fundamental challenge in computational biology. Existing methods often build a network by directly connecting genes based on pairwise relationships (e.g., correlation, mutual information). While these "single-gene" approaches have been useful, they can become computationally intensive and less effective with the massive, high-dimensional datasets generated by modern technologies like RNA-sequencing. Furthermore, these methods can struggle to capture complex regulatory logic and may produce networks that are difficult to interpret biologically. This paper addresses these limitations by proposing a "module-based" approach that first groups co-expressed genes into modules and then infers regulatory relationships at the module level, aiming for more efficient and insightful GRN reconstruction.

Methods

The proposed method consists of two main phases:

  • Phase I: Module Generation: This phase iteratively identifies groups of co-expressed genes (modules) and associates them with a small set of regulatory genes. It employs a process inspired by k-means clustering, but with a key difference: instead of using centroids, it uses "eigengenes" (the first principal component of a module's expression profiles) as representatives. These eigengenes are then modeled as a linear combination of regulator gene expressions. The authors explore several methods for this linking step, including LASSO and a novel application of Variational Bayes for Sparse Regression (VBSR). To ensure robustness and explore the solution space, this phase is bootstrapped multiple times.
  • Phase II: Bipartite Graph Generation: Once modules and their associated regulators are identified, bipartite graphs are constructed for each module. These graphs represent the relationships between individual target genes within a module and their identified regulators. This phase also uses methods like LASSO and VBSR to determine the edges of these bipartite graphs.

The paper compares this module-based approach with traditional "single-gene network" approaches where bipartite graphs are built directly from the initial expression data without prior module identification.

Datasets

The study utilizes the following datasets:

  • Simulated Data: Expression data is simulated to create a controlled environment for evaluating the performance of different methods in recovering known underlying modules and regulatory relationships.
  • TCGA (The Cancer Genome Atlas) Project Data: Specifically, the ovarian cancer (TCGA-OV) and head and neck squamous carcinoma (TCGA-HNSC) datasets are used. These datasets contain genome-wide expression data for thousands of genes across numerous cancer samples. The authors filter these datasets to retain the 10,000 most varying genes for computational efficiency. Regulatory genes are identified from the HPRD database.

Key Insights and Ideas

  • Modules as a Better Unit of Analysis: Grouping co-expressed genes into modules before inferring regulatory relationships appears to be a more efficient and biologically meaningful strategy than analyzing genes individually.
  • Variational Bayes for Module Linking Shines: The application of Variational Bayes for Sparse Regression (VBSR) to link module eigengenes with regulators consistently outperforms other methods, including LASSO, in terms of accuracy and biological relevance.
  • Bootstrapping Enhances Discovery: The use of bootstrapping in module generation is crucial. It allows for the discovery of new enriched elements that might be missed with a single run, suggesting that the heuristic nature of the method benefits from exploring multiple solutions.
  • Module-Based Networks are Sparsest and More Informative: The module-based approach, particularly with VBSR, generates sparser networks with a better ratio of target genes to regulators, leading to a higher rate of biologically interpretable enriched molecular pathways.
  • Single-Gene Networks Can Be Less Informative: Traditional single-gene network approaches, while simpler, often result in less informative networks and can struggle with the scale and complexity of modern genomic datasets.

Key Results

  • VBSR Outperforms: The module-based approach using VBSR for linking eigengenes to regulators significantly outperforms other methods (LASSO, LM) in terms of both the accuracy of regulator-target gene estimation and the biological relevance of the resulting networks, as measured by gene set enrichment analysis.
  • Improved Network Topology: VBSR-based module networks exhibit a more desirable scale-free topology and are sparser, with a better balance between regulators and target genes compared to other methods.
  • Higher Enrichment Rate: Networks generated by the module-based VBSR approach show a significantly higher rate of enriched molecular pathways, indicating greater biological interpretability.
  • Performance on Simulated Data: On simulated data, VBSR demonstrates a better trade-off between true positives and false positives, and estimates regulator effects more accurately than LASSO.
  • Clustering Performance: While VBSR excels in regulator-target gene estimation, LASSO shows slightly better performance in terms of clustering metrics (Variation of Information and Adjusted Rand Index) when module assignments are unknown, though VBSR's overall network inference is superior.

Limitations and Future Work

  • Computational Cost of Bootstrapping: While beneficial, bootstrapping adds to the computational cost. The paper notes that for some methods (like LM), the benefit of bootstrapping diminishes after a few iterations, suggesting potential optimizations.
  • Robustness of LASSO: The paper observes that LASSO's performance can be less consistent across bootstraps compared to VBSR, potentially due to its reliance on cross-validation for parameter selection.
  • Future Work: The authors suggest that further exploration of different bootstrapping strategies and initialization methods could lead to even more robust and comprehensive network inference. They also highlight the potential for integrating other omics data types to further enhance GRN reconstruction.

Summary updated

Abstract

Abstract Gene regulatory networks describe the regulatory relationships among genes, and developing methods for reverse engineering these networks are an ongoing challenge in computational biology. The majority of the initially proposed methods for gene regulatory network discovery create a network of genes and then mine it in order to uncover previously unknown regulatory processes. More recent approaches have focused on inferring modules of co-regulated genes, linking these modules with regulator genes and then mining them to discover new molecular biology. In this work we analyze module-based network approaches to build gene regulatory networks, and compare their performance to the well-established single gene network approaches. In particular, we focus on the problem of linking genes with known regulatory genes. First, modules are created iteratively using a regression approach that links co-expressed genes with few regulatory genes. After the modules are built, we create bipartite graphs to identify a set of target genes for each regulatory gene. We analyze several methods for uncovering these modules and show that a variational Bayes approach achieves significant improvement with respect to previously used methods for module creation on both simulated and real data. We also perform a topological and gene set enrichment analysis and compare several module-based approaches to single gene network approaches where a graph is built from the gene expression profiles without clustering genes in modules. We show that the module-based approach with variational Bayes outperforms all other methods and creates regulatory networks with a significantly higher rate of enriched molecular pathways. The code is written in R and can be downloaded from https://github.com/mikelhernaez/linker .
BibTeX
@misc{hernaez2018comparison,
  author = {Mikel Hernáez and Olivier Gevaert},
  title = {{Comparison of single gene and module-based methods for modeling gene regulatory networks}},
  year = {2018},
  doi = {10.1101/307884},
  howpublished = {bioRxiv},
}
Figure