Fig 1.
Systematic experimental perturbation of cells leads to predictive network models.
We perturbed cells with combinations of targeted drugs and measured the time-resolved cellular response (step 1). These measurements were used as input to derive network models of the response to arbitrary combinatorial perturbations (step 2). Using these models, we identified drug combination targets that optimally reduce cell growth and increase apoptosis in a melanoma cell line (step 3).
Fig 2.
Phenotypic and proteomic response to single and drug combinations.
The melanoma cell line A2058 was subjected to 54 drug combinations and the response of 124 (phospho-)proteins and two phenotypes (cell number and apoptosis) was measured at eight time points from 10 minutes to 67 hours. These data were used as input for model inference and prediction. The most informative data involved a subset of the proteins and phospho-proteins (black lines) that underwent the largest changes upon perturbation (AKT-pS473/474, ERK1/2-pT202/T204, cJUN-pS73, and p21) as well as apoptosis (blue line). The remaining 120 proteins and phospho-proteins together with cell counts had a less pronounced response to perturbation. Temporal response (vertical axis) is defined as log2(xperturbed(t)/xunperturbed(t)) where x(t) are concentrations or counts as in Eqs 1 and 2 and DMSO is the unperturbed control.
Fig 3.
Model selection and error estimation.
Top left: The full dataset was divided into subsets: (i) a training dataset (gray) that contained single drug control measurements (DMSO), single drugs in low (first row) and high dose (diagonal, two times low dose), (ii) a validation dataset (green) was used to estimate the optimal regularization parameter λ*, and (iii) a test dataset (blue) was used to estimate model performance. Top right: calculated values for the Bayesian Information Criterion (BIC, gray) and number of non-zero interactions (magenta) as a function of the regularization parameter λ. Bottom left: The residual sum of squares on the validation dataset was used to identify the optimal regularization parameter λ*. The best predictive model was obtained for λ* = 3 according to lowest BIC and minimal error on the validation dataset. Error bars indicate the standard deviation from 10 independent runs. Bottom right: Agreement of measured and predicted protein and phospho-protein (dots) and phenotype levels (triangles) on the test dataset. The Pearson correlation coefficient on left-out data for the combined set of molecular and phenotype nodes is 0.54, and 0.79 for phenotypic nodes alone. The mean RSS for the combined phenotypic and molecular nodes is 0.181, and 0.118 for the phenotypic nodes alone.
Fig 4.
Model-inferred effect of drugs on proteins and phospho-proteins.
The effect of drug treatment on (phospho-)protein levels is captured as edges between drugs (colored circles) and proteins (gray circles) in the model (represented by the drug–(phospo-)protein interaction strength dil from Eq 2). Some of the edges are well known (e.g., MEKi inhibits ERK1/2-pT202/T204) and some appear to be novel or indirect (e.g., inhibitory effect of PKCi on CREB-pS133). Based on the distribution of edge values over the 101 network models, only the strongest drug–protein edges are displayed for visualization purposes (85th percentile for positive/activating interactions and 15th percentile for negative/inhibiting interactions, by absolute value).
Fig 5.
Model predicts the effect of combination perturbations and suggests optimal inhibitor combinations.
The top 20 × 20 predictions of pairwise inhibition of molecular nodes (i.e., proteins and phospho-proteins) that decrease cell growth (bottom left, blue) and increase apoptosis (bottom right, red). Cell growth and apoptosis were computed for each target combination. These values were log2-transformed, normalized to the unperturbed steady state, and the average value over 101 network model predictions is presented (diagonal represents predictions for inhibition of single targets). Combinations nominated for drug testing are highlighted by dark-rimmed squares. For complete heatmaps of all tested predictions, see S8 and S9 Figs.
Fig 6.
Experimental testing of model predictions using single and pairwise drug combinations.
Several single and pairwise perturbations that were predicted to have a differing phenotypic effects were experimentally tested in the melanoma cell line A2058 (highlighted in Fig 5). Cells were perturbed with drugs that target nodes in the computational model (top). During perturbation, cells were subjected to live-cell imaging, the resulting images were segmented, and cell count was quantified, normalized relative to no-drug control and log2-transformed (see growth curves in middle panel). Agreement between model prediction (log2-transformed, normalized and averaged growth) and experiment is summarized (bottom panel).