Fig. 1
From: Dynamic interaction network inference from longitudinal microbiome data
Schematic diagram illustrating the whole computational pipeline proposed in this work. Figure shows microbial taxa Gammaproteobacteria at each step in the pipeline from a set of five representative individual samples (subjects 1, 5, 10, 32, and 48) of the gut data set. a Input is raw relative abundance values for each sample measured at (potentially) non-uniform intervals even within the same subject. b Cubic B-spline curve for each individual sample. Sample corresponding to subject 1 (dark blue) contains less than pre-defined threshold for measured time points, thus, removed from further analysis. The remaining smoothed curves enable principled estimation of unobserved time points and interpolation at uniform intervals. c Temporal alignment of each individual sample against a selected reference sample (subject 48 shown in orange). d Post-alignment filtering of samples with alignment error higher than a pre-defined threshold. Sample corresponding to subject 5 (grey) discarded. e Learning a dynamic Bayesian network (DBN) structure and parameters. Let nodes (T1,T2,T3,T4) represent microbial taxa and (C1,C2,C3) represent clinical factors shown as circles and diamonds, respectively. Figure shows two consecutive time slices ti and ti+1, where dotted lines connect nodes from the same time slice referred to as intra edges, and solid lines connect nodes between time slices referred to as inter edges. Biological relationships are inferred from edge parameters in the learned DBN which can be positive (green) or negative (red). f Original and predicted relative abundance across four gut taxa for subject 48 at sampling rate of 1 day. Performance is evaluated by average mean absolute error (MAE) between original and predicted abundance values (MAE =0.011)