A Network-based Approach for Inferring Thresholds in Co-expression Networks

Gene co-expression networks (GCNs) specify binary relationships between genes and are of biological interest because significant network relationships suggest that two co-expressed genes rise and fall together across different cellular conditions. GCNs are built by (i) calculating a co-expression measure between each pair of genes and (ii) selecting a significance threshold to remove spurious relationships among genes. This paper introduces a threshold criterion based on the underlying topology of the network. More specifically, the criterion considers both the rate at which isolated nodes are added to the network and the density of its components when the threshold varies. In addition to Pearson's correlation measure, the biweight midcorrelation, the distance correlation, and the maximal information coefficient are used to build different GCNs from the same data and showcase the advantages of the proposed approach. Finally, a case study presents a comparison of the predictive performance of the different networks when trying to predict gene functional annotations using hierarchical multi-label classification.