Graph inference enhancement with clustering: Application to Gene Regulatory Network reconstruction

The obtention of representative graphs is a key problem in an increasing number of fields, such as computer graphics, social sciences, and biology to name a few. Due to the large number of possible solutions from the available amount of data, building meaningful graphs is often challenging. Nonetheless, enforcing a priori on the graph structure, such as a modularity, may reduce the underdetermination in the underlying problem. In this work, we introduce such a methodology in the context of Gene Regulatory Network inference. These networks are useful to visualize gene interactions occurring in living organisms: some genes regulate the expression of others, structuring the network into modules where they play a central role. Our approach consists in jointly inferring the graph and performing a clustering using the graph-Laplacian-based random walker algorithm. We validate our approach on the DREAM4 dataset, showing significant improvement over state-of-the-art GRN inference methods.