Table 1.
Context-specific interactomes and the datasets used to reverse engineer them.
Fig 1.
OncoLead: Network-based protein activity inference.
(A) Drug perturbation induced genome-wide transcriptional changes are interpreted, based on multiple networks including ARACNE network, CHEA network, STRING network, and Gene knock-down (KD) network, with the VIPER algorithm, to infer changes in the activities of the regulatory proteins. The resulting four different protein activity matrixes were integrated into a single final protein activity matrix. In this way, VIPER analysis transforms drug-perturbation gene expression signatures into unbiased genome-wide regulator protein activity representations of CMoA. The left part is a simple illustration of how VIPER algorithm works based on the ARACNE network. First, ARACNE reverse engineers context-specific regulatory networks by leveraging a large collection of gene expression profiles (N > 100) from the same cellular context. Then, regulator’s activity is inferred by computing the enrichment of the genes in its regulon (from ARACNE) in every drug treatment signature sorted from the most over-expressed (colored in orange) to the most under-expressed (colored in green) genes. When there is positive or negative enrichment, the regulator is up-regulated or down-regulated (colored in red/blue). Regulator’s activity is represented by the normalized enrichment score. (B) IRS score decreases when progressively degrading the networks for MCF7 drug signatures. (C) Relative representation of how accurate each interactome is as a model for the transcriptional regulation in each of the three cell lines MCF7, PC3, and HL60 in CMAP database. Shown is the IRS in relative units for TRs inferred by OncoLead on each interactome (x-axis) / GES combination (see Methods for details). Percent IRS scores were obtained by dividing each specific IRS score by the largest score obtained across the three interactomes used in the analysis. (D) Distribution of the significant TRs inferred by OncoLead when adding increasing ratios of random noises to the Irinotecan signature in MCF7 cell line.
Fig 2.
CMAP and LINCS dataset analysis.
(A—B) Analysis for drugs targeting 75 TRs using either DTPA or DTGE for CMAP MCF7 datasets. For each TR with known inhibitors in the MCF7 dataset, we performed gene set enrichment analysis to test whether its DTPA or DTGE for its known inhibitors are significantly more inactivated or repressed compared to all other compound’s profiles (A) or to all other proteins (B) and obtained p-values from each test. Then we plotted the distributions of the–log10 p-values for DTPA (x-axis) versus DTGE (y-axis). Each triangle represents a TR. A vertical and a horizontal line were drawn at p-value equals 0.05 for DTPA and DTGE, respectively, which divide the plot into four parts: green, blue, red, and grey. (C) Enrichment analysis of the drug samples similar to TR silencing profiles on the vector of all drug samples in the same cell line sorted based on their inferred TR activity from LINCS data set. Each bar represents a cell line. Green color shows the number of TRs with significant enrichment (NES > 1.96; p < 0.05) which indicates the correlation between OncoLead CMoA inference and shRNA mediated TR silencing. Grey color shows the number of TRs without significant enrichment (NES< 1.96; p > 0.05). (D) OncoLead-inferred ESR1 activity changes (blue) and the differential ESR1 expression (red) upon letrozole treatment in vivo. Signatures are obtained by comparing different time points in responsive and non-responsive patients. The values shown in the figures are Z-scores based on p-value. The black dotted line represents Z-score = 1.96 and p-value = 0.05.
Fig 3.
Validation of predicted novel modulators for MYC and STAT3.
(A) Effect of computationally predicted drug modulators on MYC activity in luciferase assay. The compounds were tested by MYC luciferase reporter assay in MCF7 cell line. Concentrations for each compound are shown in S4 Table. The reporter activity for each compound was normalized to cell titer and compared to DMSO control. We compared drug induced activity vs. DMSO control using t-test. (p< 0.01: ***; p < 0.05: **; p<0.1: *) (B) STAT3 luciferase reporter assay results in MCF7 (blue) and SNB19 (red). The compounds were tested by STAT3 luciferase reporter assay in MCF7 and SNB19 cell line. Each compound was tested at 10uM. The activity for each drug is computed using the reporter activity normalized to cell titer and compared to DMSO control. We compared drug induced activity vs. DMSO control using t-test. (p< 0.01: ***; p < 0.05: **; p<0.1: *).
Fig 4.
Comparison of CMoA, GES and GI50 profile similarities.
(A) Venn diagram of the top 206 similar compound pairs (top 2.5%) using DTPA, DTGE and GI50 sensitivity profiles. (B) top: Enrichment of the top 206 pairs based on DTPA similarity in the vector of 8256 (129*128/2) compound pairs ranked by DTGE similarity, and vice versa; middle: Enrichment of the top 206 pairs based on DTPA similarity in the vector of 8256 (129*128/2) compound pairs ranked by GI50 correlation, and vice versa; bottom: Enrichment plot of the top 206 pairs using DTGE similarity in the vector of 8256 (129*128/2) compound pairs ranked by GI50 correlation, and vice versa.