Structure in Biological Networks: Comparative Analyses and Novel Interpretations

Abstract

The complexity and size of biological systems often necessitates network-based analyses; however, tools from network science, such as community detection, have seldom been well-characterized for such systems. In this work, this characterization was conducted for two types of biological systems: metabolic networks and functional human brain networks. For the former, community detection methods of different mathematical bases were applied to genome scale metabolic models of E. coli and K. pneumoniae. Communities were detected for both pruned and unaltered versions of two graph representations of the metabolic networks: the metabolite graph and the reaction adjacency graph. A comprehensive comparative analysis revealed three primary insights: 1. quantitative metrics which summarize community structure, such as modularity, do not reflect community membership, which varies greatly for communities detected via different methods; 2. these same metrics can be consistent across methods, yet vary with network representation, pruning method, and, in some cases, organism; and 3. probabilistic methods show promise for functional module detection, unlike methods of other formulations. These insights were used to inform an investigation of how network structure relates to antibiotic resistance. Statistically significant correlations with notable method/metric/representation dependence were identified, exemplifying not only the capacity for community detection methods to reveal fundamental relationships between network structure and clinically relevant phenotypes, but also the importance of applying such methods in an informed fashion. For the latter, similarly disparate community detection methods were applied to both real, statistically thresholded connectivity matrices representing the functional brain networks of over 5,000 adolescent subjects and data-driven synthetic brain networks. Comparative analyses revealed that statistically significant relationships between demographic/behavioral metadata and quantitative metrics for detected partitions vary by method. Moreover, in the case of synthetic brain networks, partitions detected by probabilistic methods were most similar to ground truth partitions. These probabilistic methods were subsequently shown to be promising avenues for characterizing hierarchical organization in the brain, investigating the evolution of said hierarchy with neural maturation, and reconstructing a consensus network based on multiple subjects. In summary, comprehensive comparisons of community detection methods were conducted for two distinct biological systems. Method dependence was revealed in both cases, and these methods were subsequently applied in informed, novel ways, elucidating key relationships between network structure and biological function.