Fig 1.
A) Model fit. The VAR model is fit on both the original and a permuted data set (blue arrows indicate shuffling each gene’s expression independently across time). Based on the null distribution of coefficients, a threshold is chosen to control the edge FDR at ≤ 0.05. B) Stability selection. From the original data, 1000 bootstrap samples are generated. For each sample, a network is inferred as in A. Each edge’s selection frequency across the bootstrapped networks is computed. C) Statistical significance. For both the original and permuted data, a selection frequency distribution is generated for stability selection as in B. Edges are thresholded to control the stability FDR at ≤ 0.2. See S1 Fig for an overview of network inference methods.
Fig 2.
Algorithm performance on the DREAM community benchmark.
A) AUPR scores from 24 methods, averaged across the five DREAM networks. B) AUROC scores from 19 methods, averaged across the five DREAM networks. Arrows indicate our methods. Stars indicate methods that we ran in-house; results were consistent with reported results. The bars reach one standard deviation from the average as calculated across the five DREAM networks; no bar indicates the standard deviation was not reported. See also S1–S5 Tables.
Fig 3.
Causal network inferred from glucocorticoid receptor data.
A) Causal network clustered by gene type. Edge color indicates the type of the causal gene: red edge indicates an immune causal edge, blue edge indicates a metabolic causal edge, purple edge (both) indicates an immune and metabolic causal edge, and tan edge indicates a neither immune nor metabolic other causal edge. B) Significance thresholding for edges, based on the null distribution of selection frequencies. C) Out-degree distribution of network. For clarity, several high out-degree values with low frequencies are not shown. D) In-degree distribution of network. E) Quantile-quantile (Q-Q) plot of in-degree distribution against normal quantiles. The in-degrees have a heavier left tail and lighter right tail than the normal distribution. F) Enrichment of gene classes among network causal genes, measured by odds ratio. G) Enrichment of edge classes among network edges, measured by odds ratio. See also S6 Table.
Fig 4.
Conditional Granger causality reveals opposite sign of relationship KRT6A → NKAIN4.
A) Time series and B) scatter plot of expression values from KRT6A and NKAIN4. C) Time series and D) scatter plot of expression values from KRT6A and residual expression values from NKAIN4 after controlling for the effects of other covariates in NKAIN4. Each y-axis tick in A and C indicates 0.1 unit-variance standardized ln(TPM), where TPM is Transcripts Per kilobase Million. The grey line marks zero-centered expression. B and D axes are in units of ln(TPM).
Fig 5.
Time-series profiles of experimentally validated causal interactions across gene classes.
For each gene pair, their profiles were from either the original exposure data set or the unperturbed data set. The effects of all covariates beside the causal gene were controlled in the effect gene values to show the conditional Granger-causal relationship. Colors encode gene classes: pink shows immune genes, dark blue/gray shows metabolic genes, teal shows TFs, and brown/tan shows other genes. Darker colors show causal genes and lighter colors show effect genes. The grey line marks zero-centered expression. Each y-axis tick indicates 0.1 unit-variance standardized ln(TPM). See also S7 Table and S2 Fig.
Fig 6.
Validation of inferred network on overexpression data.
A-B) Regression of one-hot encoding of positive (negative for B) edges as the predictor against the VAR model edge coefficient from the overexpression data as the response. A 1 indicates that an edge had a positive (in A) or negative (in B) coefficient in the original inferred network (FDR ≤ 0.2). C) For the 123 causal edges from TFCP2L1, regression of edge sign as the predictor against the VAR model edge coefficient from TFCP2L1 overexpression data as the response.
Fig 7.
Network edge validation using known cis- elements from GTEx v6 lung cis-eQTLs.
A) Enrichment of trans associations in primary lung tissue among p-values from edges inferred by BETS compared to p-values from permutations. B) Quantile-quantile plot of validated edges shows signal enrichment in lung samples when compared to signals from four other tissues in the GTEx v6 study. C) SNPs associated with inferred gene pairs. Genotype-phenotype plots corresponding to the cis-effect (left column), correlation in the GTEx v6 data between cause (y-axis) and effect (x-axis) gene pairs (right column).