Network reconstruction with local partial correlation: a comparative evaluation

TL;DR

This paper compares three network reconstruction methods for Gaussian Graphical Models, showing that Local Partial Correlation and Graphical Ridge often outperform Graphical Lasso in simulated and real gene expression data.

Contribution

It provides a comparative evaluation of LPC, GLasso, and GGMridge methods using simulated and real high-dimensional data, highlighting their relative performance.

Findings

LPC outperforms GLasso in most simulated scenarios.

GGMridge yields better ROC curves than LPC and GLasso.

LPC and GGMridge have similar performance on real data.

Abstract

Over the past decade, various methods have been proposed for the reconstruction of networks modeled as Gaussian Graphical Models. In this work, we analyzed three different approaches: the Graphical Lasso (GLasso), the Graphical Ridge (GGMridge), and the Local Partial Correlation (LPC). For the evaluation of the methods, we used high dimensional data generated from simulated random graphs (Erd\"os-R\'enyi, Barab\'asi-Albert, Watts-Strogatz). The performance was assessed through the Receiver Operating Characteristic (ROC) curve. In addition, the methods were used to reconstruct the co-expression network for differentially expressed genes in human cervical cancer data. The LPC method outperformed the GLasso in most simulated cases. The GGMridge produced better ROC curves then both the other methods. Finally, LPC and GGMridge obtained similar outcomes in real data studies.