Revisiting the General Concept of Network Centralities: A Propose for Centrality Analysis in Network Science

Background: In network science, although different types of centrality measures have been introduced to determine important nodes of networks, a consensus pipeline to select and implement the best-tailored measure for each complex network is still an open field. In the present study, we examine the node centrality profiles of protein-protein interaction networks (PPINs) in order to detect which measure is succeeding to predict influential proteins. We study and demonstrate the effects of inherent topological features and network reconstruction approaches on the centrality measure values. Results: PPINs were used to compare a large number of well-known centrality measures. Unsupervised machine learning approaches, including principal component analysis (PCA) and clustering methods, were applied to find out how these measures are similar in terms of characterizing and assorting network influential constituents. We found that the principle components of the network centralities and the contribution level of them demonstrated a network-dependent significance of these measures. We show that some centralities namely Latora, Decay, Lin, Freeman, Diffusion, Residual and Average had a high level of information in comparison with other measures in all PPINs. Finally, using clustering analysis, we restated that the determination of important nodes within a network depends on its topology. Conclusions: Using PCA and identifying the contribution proportion of the variables, i.e., centrality measures in principal components, is a prerequisite step of network analysis in order to infer any functional consequences, e.g., the essentiality of a node. Our conclusion is based on the signal and noise modeling using PCA and the similarity distance between clusters. Also, an interesting strong correlation between silhouette criterion and contribution value was found which corroborates our results.