Network fingerprint: a knowledge-based characterization of biomedical networks

High-throughput experimental technologies enabled biological studies be performed at the network scale in recent years. Various molecular networks (protein-protein interaction, metabolic, signaling and transcriptional regulatory networks) have become basic objects in biomedical studies1. Powerful computational tools are therefore needed to help researchers gain insight into the meaning of biological networks. Several approaches have been proposed to decipher these complex objects, such as identifying topological important nodes, finding structural motifs2, and detecting communities3,4. However, these approaches may involve too many new mathematical concepts, and cannot provide a familiar “language” for biomedical researchers that clearly indicate biological functions or diseases. Unlike researches on other kinds of complex networks, in biomedical studies, a number of basic networks with clear functions, referred to as “pathways” or “modules”, provide a good basis for knowledge-based exploration of biomedical networks. For example, the well-known KEGG database now contains a large number of manually-created pathways representing current knowledge of molecular interactions and reaction networks for several processes, including metabolism, genetic information processing, and cellular processes5. This situation inspired us to adopt the comparison and decomposition strategies for molecular network understanding. In this study, we introduce a framework to characterize a biomedical network using a series of its similarities (both structural and functional) to a set of basic networks, which we call “network fingerprint”. Given a set of basic networks , a biomedical network G can be characterized by a network fingerprint, where si = sim (G, Pi) is the functional similarity between G and Pi. It is important to choose proper basic networks set. Here, we use the well-studied KEGG signaling pathways as the set of basic networks. Each of these pathways represents certain cellular process, which is familiar to biomedical researchers. Moreover, the KEGG signaling pathways are relatively independent with each other, and have proper network size. To compute the network fingerprint, we presented an algorithm to measure the functional and structural similarity between G and Pi based on the gene ontology (GO) and affinity propagation (AP) clustering algorithm6, and the similarity score is normalized by the random simulation procedure (details are shown in the Supplementary files). The network contains a sub-network similar to a certain basic network have a corresponding high score. In fact, there are already several methods to measure the similarity of two networks in computer science. These methods are almost based on topological structure of networks but ignore the network node property. However, the biological function of the proteins is an important factor when compare two biomedical networks. Differently, our method takes both the topological structure and the function of the proteins in the network into consideration. Based on this approach, we provide new insights into the space of disease networks as well as the relationships between diseases and signaling pathways.