Sparse and Compositionally Robust Inference of Microbial Ecological Networks
Fig 2
Workflow of the SPIEC-EASI pipeline.
The SPIEC-EASI pipeline consists of two independent parts for a) synthetic data generation and b) network inference. a) Synthetic data generation requires an OTU count table and a user-selected network topology. Internally, the parameters of a statistical distribution (the zero-inflated Negative binomial model is suggested) are fit to the OTU marginals of the real data, and are combined with the randomly-generated network in the Normal to Anything (NORTA) approach to generate correlated count data. b) Network inference proceeds in three stages on synthetic or real OTU count data: First, data is pre-procssed and centered log-ratio (CLR) transformed to ensure compositional robustness. Next, the user selects one of two graphical model inference procedures: 1) Neighborhood selection (the MB method) or 2) inverse covariance selection (the glasso method). SPIEC-EASI network inference assumes that the underlying network is sparse. We infer the correct model sparseness by the Stability Approach to Regularization Selection (StARS), which involves random subsampling of the dataset to find a network with low variability in the selected set of edges. SPIEC-EASI outputs include an ecological network (from the non-zero entries of the inverse covariance network) and an invertible covariance matrix. If the network was inferred from synthetic data, it can be compared with the input network to assess inference quality.