Fig 1.
Workflow of loop identification GSE raw datasets were processed using Metageo to obtain gene expression data with corresponding gene names.
Fig 2.
Schematic demonstrating use of symbolic aggregate approximation (SAX) in identifying candidate phase-shifted gene pairs.
(A) Selected large-ranged genes were converted from numerical log2 form to SAX form. For each gene, we then computed a search pattern by shifting the SAX form by T/4 time points (T = total number of time points sampled) (B) Determining phase-shifted gene pairs based on SAX search pattern match; Genes 1 and 2 match the pattern while Genes 1 and 3 do not. (C) Genes that match the search pattern trace a loop (left) while mismatched genes do not (right).
Fig 3.
IL1A-TNIP3 loop in human monocytes can predict perturbation stage in test samples.
(A) Gene pairs tracing loops identified using training data from the human monocyte dataset (GSE47122). Each loop represents the space traced by a gene pair using median gene expression across individuals. (B) Gene expression profile of IL1A and TNIP3 highlighting phase shift. (C) IL1A-TNIP3 loop. Circles represent data points for all individuals and time points in the training data. (D) Polar plot derived from IL1A-TNIP3 loop. Points are colored based on an ordinal time scale to clearly distinguish between points sampled at different times (0 on this ordinal scale corresponds to time point 0 hrs, 1 to 2 hrs, 2 to 2.5 hrs, 3 to 3 hrs, 4 to 3.5 hrs, 5 to 4 hrs, 6 to 14 hrs, 7 to 24 hrs, and 8 to 48 hrs). Distinct time points can be seen to occupy distinct regions on the plot. (E) Angle, derived from polar transformation of the loop, is positively linearly correlated with time on the x-axis (Pearson’s ρ = 0.98) (F) Prediction of perturbation stage for 34 test samples, computed using k-nearest neighbors, shows an accuracy of 94% with a strong linear relationship between predicted and actual time (R2 = 0.99).
Fig 4.
Experimental validation of IL1A-TNIP3 loop.
(A) Experimental confirmation of IL1A-TNIP3 loop in a representative human donor among a total of three donors. Circles represent technical and biological replicates. Colors represent time points, labeled next to the figure; distinct colors represent distinct exposures to given stimuli described in the Results. Different doses of LPS were used (5ng/ml, 50ng/ml and 500ng/ml) but a dose-dependent response was not observed. (B) Gene expression of IL1A (red, left) and TNIP3 (blue, right) over time highlighting phase-shift. Increasing color intensities correlate with increasing doses of LPS.
Fig 5.
Identifying CDC20-IFI44L loop and predicting perturbation stage in YF17D-vaccinated individuals.
(A) Timeline of gene expression shows that CDC20 peaks at day 3 whereas IFI44L peaks at day 7 for training data in the Montreal cohort (n = 11) (GSE13699). (B) CDC20-IFI44L loop. Circles represent data points for all individuals and time points in the point cloud. Two data points have been removed for better visualization; all data points are shown in S6 Fig. (C) Angle derived from polar transformation of the CDC20-IFI44L loop is positively linearly correlated with and time (Pearson’s ρ = 0.91) (D) Confusion matrix, in the form of a heatmap, showing prediction accuracy for stage of infection in the withheld samples in the Montreal cohort (n = 7) (GSE13699). A stage prediction accuracy of 83% was achieved. (E) Confusion matrix, in the form of a heatmap, showing prediction accuracy for stage of infection in the Lausanne cohort (n = 11) (GSE13699). A stage prediction accuracy of 73% was achieved. (F) Confusion matrix, in the form of a heatmap, showing prediction accuracy for stage of infection in the Emory cohort (n = 25) (GSE13485); a prediction accuracy of 65% was achieved.