Comparison of deep and shallow graph representation learning algorithms for detecting modules in molecular networks

Graphs provide ubiquitous powerful representations of interacting components in many complex systems. In molecular biology, graphs are used to represent interactions between molecular components; for example, protein-protein interactions represent interactions between proteins, while gene regulatory networks describe interactions between transcription factor proteins and genes. A common property of graphs for complex systems, including molecular networks, is the presence of community structures or modules, defined by groups of densely interacting nodes. In protein-protein interaction networks, such modules often correspond to key biological processes needed for cells to accomplish different tasks [1]. Disruptions to module connectivity have been implicated in many diseases [13]. Therefore, the accurate detection of these modules is a key problem in the analysis of molecular networks [14]. Graph representation learning [15], which aims to embed graph nodes in a d-dimensional space (Fig 1), offers a powerful framework for performing machine learning tasks on graphs, including clustering. Both deep and shallow graph representation algorithms exist; however, deep representation learning methods have been designed and tested on social networks, citation networks and recommendation systems. Therefore, it is unclear if deep representation learning methods are beneficial for module detection of molecular networks. In this work, we compare several state-of-the-art shallow and deep learning frameworks on both real molecular networks and synthetic benchmark networks.