Predicting viral exposure response from modeling the changes of co-expression networks using time series gene expression data

摘要

In genomics studies, time-series gene expression data [1–3] often need to be processed and analyzed. In 2016, DREAM CHALLENGES released an open challenge called ‘Respiratory Viral DREAM Challenge: Discovering dynamic molecular signatures in response to virus exposure’ (https://www.synapse.org/#!Synapse:syn5647810/wiki/399108). The aim was to develop early predictors of susceptibility and contagiousness based on expression profiles collected prior to and at early time points following viral exposure. Some work reported the differences of transcriptomics [4–6] in the host response between symptomatic and asymptomatic subjects exposure to respiratory viruses. Additionally, as what were done by most participants (https://www.synapse.org/#!Synapse:syn5647810/wiki/402364), some common machine learning algorithms [7] can be used if we treat the challenge as a prediction problem. The challenge results [see Additional file 1 for parts of the challenge results] demonstrate that the prediction performance significantly depends on the participants’ models. However, we need to average the time series data across time or only use cross-sectional data at a time to perform ensemble learning, and the dynamic information of the time series data is lost in these approaches. Moreover, in the early stage of infection (within 24 h), there is little separation of the trajectories of genes among subjects with different clinical responses. Previous studies [8, 9] also showed that the individual responses after exposure to respiratory virus are influenced not only by the baseline immune status of the host but also by the dynamics of the early host immune response immediately following exposure. If we only consider a single gene, there is no distinct pattern in both cross-sectional and dynamic data. It is difficult to differentiate between positive and negative groups by gene expression levels at early stage. In this paper, we resort to gene sets analysis to correlate exposure response with dynamic gene expression patterns in gene sets. To consider multiple genes, some methods have been proposed to infer the relationship between genes. For example, the Dynamic Bayesian Network (DBN) was used to establish the dynamic regulatory network [10]. We note that a number of groups have studied time-varying dynamic Bayesian networks (TV-DBN) to model the varying network structures and reveal the dynamics of biological systems [11, 12]. The dynamic mixed membership stochastic block model (dMMSB) helps to infer the biological functions of genes through modeling the dynamic tomography of networks [13]. The review of differential network biology [14] advocated that differential network mapping at large scales may provide a deeper understanding of complex biological phenomena. The work [15] analyzed multiple differential co-expression networks based on time-course RNA-Seq data. Through Multiple Differential Modules (M-DMs), they found that dynamic modules are associated with the development of heart failure. These results in the literature suggest that considering the dynamics of networks may help us to better understand disease onset and progression. However, how to extract useful dynamic information from time-series gene expression data to build predictive model remains a challenging problem.