Comparing statistical methods for constructing large scale gene networks

Abstract

The gene regulatory network (GRN) reveals the regulatory relationships among genes and can provide a systematic understanding of molecular mechanisms underlying biological processes. The importance of computer simulations in understanding cellular processes is now widely accepted; a variety of algorithms have been developed to study these biological networks. The goal of this study is to provide a comprehensive evaluation and a practical guide to aid in choosing statistical methods for constructing large scale GRNs. Using both simulation studies and a real application in E. coli data, we compare different methods in terms of sensitivity and specificity in identifying the true connections and the hub genes, the ease of use, and computational speed. Our results show that these algorithms performed reasonably well, and each method has its own advantages: (1) GeneNet, WGCNA (Weighted Correlation Network Analysis), and ARACNE (Algorithm for the Reconstruction of Accurate Cellular Networks) performed well in constructing the global network structure; (2) GeneNet and SPACE (Sparse PArtial Correlation Estimation) performed well in identifying a few connections with high specificity.

Figure 1. Diagram depicting the network structures of each of the six network sizes used in this study.

Figure 2. The AUCs and pAUCs for 1344-gene network in simulation 1.

Left: The area under the curve using various network construction methods across various sample sizes on a network with 1344 genes; Right: The partial area under the curve for FPR0.005 for various methods.

Figure 3. Comparison of the AUC performance on detecting hub genes.

Measuring the performance of each method at detecting hub genes as measured by the Area Under the ROC Curve (AUC). Hub genes were classified as having 4 or more connections in the true network.

Figure 4. The AUCs for 1344-gene network in simulation 2 with non-normal distribution.

The area under the curve using various network construction methods across various sample sizes on a network with 1344 genes.

Figure 5. The transcriptional regulatory network for E. coli derived from the RegulonDB database.

Each red dot is a gene, and a blue line between genes indicates a connection.

Figure 6. The performance in constructing gene regulatory network in E. coli.

Left: The entire ROC curves using various network construction methods; Right: The corner of ROC with high specificity.