Fig 1.
The classical network inference pipeline.
Fig 2.
High level summary of the OneNet pipeline.
(i) bootstrap subsamples are constructed from the original abundance matrix, (ii) each inference method is applied on the bootstrap subsamples to compute edge selection frequencies using a fixed λ grid, (iii) a different λ is selected for each method to achieve the same density in all methods, (iv) edge selection frequencies are summarized and (v) thresholded to compute the consensus graph.
Table 1.
Characteristics of the network inference methods.
Fig 3.
Precision—Recall curves of each inference method for different sample sizes.
(a) 50 (b) 100 (c) 500 (d) 1000. The TPR/PPV compromise achieved for λ* corresponding to a stability of 0.9 is shown with a circle, the one achieved by a mean stability across methods of 0.9 is shown with a square. Whenever the selected λ is the same, the circle and the square are replaced with a diamond. Finally, note that COZINE relies on minimization of a BIC criteria (shown with a triangle) rather than on the resampling-based stability selection to choose the regularization parameter.
Fig 4.
Quality of both single-method and consensus network in terms of PPV/TPR assessed on 20 simulated datasets of size n = 100 samples.
Striped (resp. no-strip) boxplots show PPV (resp. TPR) values.
Fig 5.
Consensus and single-method networks inferred on the liver cirrhosis dataset, followed by CORE-clustering algorithm to identify the microbial guilds.
All graphs share the same layout, computed on the OneNet-mean network, to ease comparisons. Nodes are colored by cluster, with red always used for the cirrhotic guild in all graphs where it is (at least partially) recovered and species prevalent in the oral cavity are represented by a triangle. Methods are grouped based on the underlying inference technique (tree aggregation, graphical lasso, neighborhood selection).
Fig 6.
Detailed view of the cirrhotic guild identified in the OneNet-mean network with taxonomic information on the nodes.
All species known to be associated with chronic diseases are marked with a star (*).