Analysis of bio-molecular networks through semi-supervised graph-based learning methods

Relevant problems in the context of molecular biology and medicine can be modeled through graphs where nodes represent bio-molecular or chemical entities (e.g. genes or drugs) and edges some notion of similarity between them. In this context, semi-supervised learning methods able to exploit both the local (e.g. the neighborhood of a node) and the global characteristics of the network (e.g. its overall topology) have been applied to extract meaningful biological and medical knowledge from a biological system. In this work we summarize the main characteristics of RANKS (RAnking Nodes through Kernelized Score functions), a recently proposed semi-supervised algorithmic scheme based on local score functions embedding well-designed graph kernels, able to deal with both the local and the global features of the analyzed network. We show some successful applications of RANKS in the context of protein function prediction, gene disease association and drug repositioning problems. Moreover we present a novel secondary memory-based and "vertex-centric" version of the algorithm able to nicely scale on graphs with hundreds of thousands of nodes and tens of millions of edges, using off-the-shelf desktop computers, and we show an application to a complex multi-species protein function prediction problem.