ESNMF: Evolutionary Symmetric Nonnegative Matrix Factorization for Dissecting Dynamic Microbial Networks

Abstract

Dynamic network is drawing more and more attention due to its potential in capturing time-dependent phenomena such as online public opinion and biological system. Microbial interaction networks that model the microbial system are often dynamic, static analysis methods are difficult to obtain reliable knowledge on evolving communities. To fulfill this gap, a dynamic clustering approach based on evolutionary symmetric nonnegative matrix factorization (ESNMF) is used to analyze the microbiome time-series data. To our knowledge, this is the first attempt to extract dynamic modules across time-series microbial interaction network. ESNMF systematically integrates temporal smoothness cost into the objective function by simultaneously refining the clustering structure in the current network and minimizing the clustering deviation in successive timestamps. We apply the proposed framework on a human microbiome datasets from infants delivered vaginally and ones born via C-section. The proposed method cannot only identify the evolving modules related to certain functions of microbial communities, but also discriminate differences in two kinds of networks obtained from infants delivered vaginally and via C-section.