Learning complex protein interaction network representations

AGUSTINA DINAMARCA; DIEGO BUSTOS; LUIS G. MOYANO

Mendoza

Congreso; A2B2C 10th Meeting; 2019

Asociación Argentina de Bioinformática y Biología Computacional

BACKGROUND: Representation learning is an umbrella term for a combination of techniques that aims to efficiently map data structures into convenient latent spaces, either for dimensionality reduction or for expressing semantic content. Embedding methods to represent networks in low dimensional vector spaces are useful tools for understanding complex network structure. Here we study the properties of two embedding methods based specifically designed for networks.RESULTS: We analyze the results of applying two random-walk embedding methods (deepwalk and node2vec) to several protein-protein interaction networks (e.g., mouse, E. coli, human, among others). The low-dimensional representation (for different values of dimensionality) in Euclidian space is afterwards analyzed in d=2 space through dimensionality reduction with t-SNE. Finally, the result is inspected for clusters using DBSCAN. We find consistent clustering partitions across the methods' parameter space. These clusters present very low resemblance (as expressed by Jaccard similarity index) to the ones obtained via traditional network community detection methods.CONCLUSIONS: Our results suggest there is structural information contained in the interactomes accessible through these class of embedding methods, and not apparent in clusters obtained via traditional network methods.