Comparative analysis of false discovery rate methods in constructing metabolic association networks.

Gaussian graphical model (GGM)-based method, a key approach to reverse engineering biological networks, uses partial correlation to measure conditional dependence between two variables by controlling the contribution from other variables. After estimating partial correlation coefficients, one of the most critical processes in network construction is to control the false discovery rate (FDR) to assess the significant associations among variables. Various FDR methods have been proposed mainly for biomarker discovery, but it still remains unclear which FDR method performs better for network construction. Furthermore, there is no study to see the effect of the network structure on network construction. We selected the six FDR methods, the linear step-up procedure (BH95), the adaptive linear step-up procedure (BH00), Efron's local FDR (LFDR), Benjamini-Yekutieli's step-up procedure (BY01), Storey's q-value procedure (Storey01), and Storey-Taylor-Siegmund's adaptive step-up procedure (STS04), to evaluate their performances on network construction. We further considered two network structures, random and scale-free networks, to investigate their influence on network construction. Both simulated data and real experimental data suggest that STS04 provides the highest true positive rate (TPR) or F1 score, while BY01 has the highest positive predictive value (PPV) in network construction. In addition, no significant effect of the network structure is found on FDR methods.