Recursive indirect-paths modularity (RIP-M) for detecting community structure in RNA-seq co-expression networks

Clusters of genes in co-expression networks are commonly used as functional units for gene set enrichment detection and increasingly as features (attribute construction) for statistical inference and sample classification. One of the practical challenges of clustering for these purposes is to identify an optimal partition of the network where the individual clusters are neither too large, prohibiting interpretation, nor too small, precluding general inference. Newman Modularity is a spectral clustering algorithm that automatically finds the number of clusters, but for many biological networks the cluster sizes are suboptimal. In this work, we generalize Newman Modularity to incorporate information from indirect paths in RNA-Seq co-expression networks. We implement a merge-and-split algorithm that allows the user to constrain the range of cluster sizes: Large enough to capture genes in relevant pathways, yet small enough to resolve distinct functions. We investigate the properties of our recursive indirect-pathways modularity (RIP-M) and compare it with other clustering methods using simulated co-expression networks and RNA-seq data from an influenza vaccine response study. RIP-M had higher cluster assignment accuracy than Newman Modularity for finding clusters in simulated co-expression networks for all scenarios, and RIP-M had comparable accuracy to Weighted Gene Correlation Network Analysis (WGCNA). RIP-M was more accurate than WGCNA for modest hard thresholds and comparable for high, while WGCNA was slightly more accurate for soft thresholds. In the vaccine study data, RIP-M and WGCNA enriched for a comparable number of immunologically relevant pathways.