Summary Summary (text) Abstract Abstract (text) MEDLINE XML PMID List

#### Send to
jQuery(document).ready( function () {
jQuery("#send_to_menu input[type='radio']").click( function () {
var selectedValue = jQuery(this).val().toLowerCase();
var selectedDiv = jQuery("#send_to_menu div." + selectedValue);
if(selectedDiv.is(":hidden")){
jQuery("#send_to_menu div.submenu:visible").slideUp();
selectedDiv.slideDown();
}
});
});
jQuery("#sendto").bind("ncbipopperclose", function(){
jQuery("#send_to_menu div.submenu:visible").css("display","none");
jQuery("#send_to_menu input[type='radio']:checked").attr("checked",false);
});

File Clipboard Collections E-mail Order My Bibliography Citation manager

Format Summary (text) Abstract (text) MEDLINE XML PMID List CSV

- 1 selected item: 26284497
Format Summary Summary (text) Abstract Abstract (text) MEDLINE XML PMID List MeSH and Other Data E-mail Subject Additional text

Generate a file for use with external citation management software.

# Perturbation biology nominates upstream-downstream drug combinations in RAF inhibitor resistant melanoma cells.

^{1}, Wang W

^{1}, Demir E

^{1}, Aksoy BA

^{1}, Jing X

^{1}, Molinelli EJ

^{1}, Babur Ö

^{1}, Bemis DL

^{1}, Onur Sumer S

^{1}, Solit DB

^{2}, Pratilas CA

^{3}, Sander C

^{1}.

### Author information

- 1
- Computational Biology Center, Memorial Sloan Kettering Cancer Center, New York, United States.
- 2
- Human Oncology and Pathogenesis Program, Memorial Sloan Kettering Cancer Center, New York, United States.
- 3
- The Sidney Kimmel Comprehensive Cancer Center at Johns Hopkins University, Baltimore, United States.

### Abstract

Resistance to targeted cancer therapies is an important clinical problem. The discovery of anti-resistance drug combinations is challenging as resistance can arise by diverse escape mechanisms. To address this challenge, we improved and applied the experimental-computational perturbation biology method. Using statistical inference, we build network models from high-throughput measurements of molecular and phenotypic responses to combinatorial targeted perturbations. The models are computationally executed to predict the effects of thousands of untested perturbations. In RAF-inhibitor resistant melanoma cells, we measured 143 proteomic/phenotypic entities under 89 perturbation conditions and predicted c-Myc as an effective therapeutic co-target with BRAF or MEK. Experiments using the BET bromodomain inhibitor JQ1 affecting the level of c-Myc protein and protein kinase inhibitors targeting the ERK pathway confirmed the prediction. In conclusion, we propose an anti-cancer strategy of co-targeting a specific upstream alteration and a general downstream point of vulnerability to prevent or overcome resistance to targeted drugs.

#### KEYWORDS:

cancer drug resistance; cell biology; cellular signaling; computational biology; drug synergy; human; melanoma; network modeling; proteomics; systems biology

### Comment in

- Perturbing resistance: a network perspective. [Pigment Cell Melanoma Res. 2016]

- PMID:
- 26284497
- PMCID:
- PMC4539601
- DOI:
- 10.7554/eLife.04640

- [Indexed for MEDLINE]

**DOI:**http://dx.doi.org/10.7554/eLife.04640.003

_{ij}). The algorithm chart depicts one round of BP-guided decimation to generate a single model solution. Consecutive runs of BP-guided decimation algorithm lead to construction of a network model solution ensemble.

**DOI:**http://dx.doi.org/10.7554/eLife.04640.004

**A**) The combinatorial perturbation matrix. The melanoma cells are perturbed with combinations of targeted drugs (see for perturbation conditions). (

**B**) The concentration changes in 138 proteomic entities (50 phospho, 88 total protein measurements) (left) and the phenotypic changes (right) in response to drug combinations with respect to the untreated conditions form an experimental ‘response map’ of the cellular system. The response map reflects the functional relations between signaling proteins and cellular processes. The two-way clustering analysis of the proteomic readouts reveals distinct proteomic response signatures for each targeted drug. Phosphoproteomic response is measured using the RPPA technology. Cell cycle progression and viability response are measured using flow cytometry and resazurin assays, respectively. The cell cycle progression phenotype is quantified based on the percentage of the cells in a cell cycle state in perturbed condition with respect to the unperturbed condition. For the phenotypic readouts, the order of the perturbation conditions is same as in (

**A**). The response values are relative to a no drug control and given as log

_{2}(perturbed/unperturbed).

**DOI:**http://dx.doi.org/10.7554/eLife.04640.008

**DOI:**http://dx.doi.org/10.7554/eLife.04640.009

**DOI:**http://dx.doi.org/10.7554/eLife.04640.010

**DOI:**http://dx.doi.org/10.7554/eLife.04640.011

**DOI:**http://dx.doi.org/10.7554/eLife.04640.012

**A**) To test the predictive power of network models, a leave-11-out cross-validation test is performed. Using the BP-guided decimation algorithm, 4000 network model solutions are inferred in the presence and absence of prior information using the partial response data. Resulting models are executed with in silico perturbations to predict the withheld conditions. Each experimental data point represents the readouts from RPPA and phenotype measurements under the corresponding perturbation conditions. Each predicted data point is obtained by averaging results from simulations with in silico perturbations over 4000 model solutions. The experimental and predicted profiles are compared to demonstrate the power of network models to predict response to combinatorial drug perturbations. (

**B**) In all conditions, network inference with prior information leads to a higher cumulative correlation coefficient (R) and significantly improved prediction quality (RAFi p = 1 × 10

^{−3}, AKTi p = 5.7 × 10

^{−3}, unpaired t-test H

_{0}: ΔX

^{with_prior}= ΔX

^{w/o_prior}, ΔX = |X

_{exp}− X

_{pred}|) between experimental and predicted responses. Plots on top row: prior information is used for network inference. Plots on bottom row: no prior information is used for network inference. Response to RAFi + {D

_{i}} (first and second column) and AKTi + {D

_{i}} (third and fourth column) is withheld from the training set and the withheld response is predicted. All responses (phenotypic + proteomic) (first and third column) and only phenotypic responses (second and fourth column) are plotted. {D

_{i}} denotes set of all drug perturbations combined with drug of interest. (See for the cross-validation calculations with all other partial data sets and for statistical validation of the predictions.)

**DOI:**http://dx.doi.org/10.7554/eLife.04640.013

**DOI:**http://dx.doi.org/10.7554/eLife.04640.014

**A**) The generation of the average model. The set of <W

_{ij}> averaged over the W

_{ij}in all models provide the average network model. The signaling processes are explained through qualitative analysis of the average model and its functional subnetworks (see for quantitative analysis). (

**B**) The average network model contains proteomic (white) and phenotypic nodes (orange) and the average signaling interactions (<Wij> > 0.2) over the model solutions. The edges between the BRAF, CRAF, TSC2, and AKTpT308 represent the cross-pathway interactions between the MAPK and PI3K/AKT pathways (see for analysis of edge distributions in the solution ensemble) (green edges: positive signed, red edges: negative signed interactions). (

**C**) Cell cycle signaling subnetwork contains the interactions between the cyclins, CDKs, and other associated molecules (e.g., p27/Kip1). RBpS807 and cyclin D1 are the hub nodes in the subnetwork and connect multiple signaling entities. (

**D**) ERK subnetwork. MEKpS217 is the critical hub in this pathway and links upstream BRAF and SRC to downstream effectors such as ERK phosphorylation. (

**E**) In the PI3K/AKT subnetwork, the SRC nodes (i.e., phosphorylation, total level, activity) are upstream of PI3K and AKT (total level, AKTpS473 and AKTp308) and the AKT nodes are the major hubs. Downstream of AKT, the pathway branches to mTOR, P70S6K and S6 phosphorylation cascade and the GSK3β phosphorylation events. A negative edge originating from mTORpS2448 and acting on AKTpS473 presumably captures the well-defined negative feedback loop in the AKT pathway (). Note that nodes tagged with ‘a’ (e.g., aBRAFV600E) are activity nodes, which couple drug perturbations to proteomic changes.

**DOI:**http://dx.doi.org/10.7554/eLife.04640.015

**DOI:**http://dx.doi.org/10.7554/eLife.04640.016

_{ij})) in the solution ensemble. The y-axis represents the frequency values for nonzero edge strengths (f(|w

_{ij}| > 0.2)) in the ensemble. The X-axis represents the interactions ordered by their frequencies, f(|w

_{ij}| > 0.2). The part of the long tail, which corresponds to the edges with frequencies less than 0.05, is not displayed. A set of well reported interactions such as the coupling of AKT activity node to phospho-AKT and phospho-MEK to phosho-MAPK (ERK) are inferred with |w

_{ij}| > 0.2 in >99% of the model solutions (left). Few examples of the rarely captured signaling interactions are given on right. Bottom: the frequency distribution of nonzero edges (|w

_{ij}| > 0.2) has a bimodal character. A set of interactions with nonzero edge values is shared by nearly all of the inferred network models (0.8 < f(|w

_{ij}| > 0.2) < 1.0). Such frequent edges form a stable network skeleton shared by majority of the solutions. On the other hand, a set of possible interactions have nonzero edge strength (w

_{ij}) values in ∼50% of the solutions. Such interactions vary among different model solutions and possibly account for the variations in system response predicted with different model solutions. The interactions in the prior model are particularly enriched in these two groups. A large fraction of potential interactions have very low (0.05 < f(|w

_{ij}| > 0.2) < 0.20) or near zero (f(|w

_{ij}| > 0.2) < 0.05) frequencies in the model ensemble.

**DOI:**http://dx.doi.org/10.7554/eLife.04640.017

_{ij})) for each possible interaction. Next, we constructed the models by assigning the edge value (W

_{ij}) for which BP-generated P(W

_{ij}) is maximum. We compared the models generated with different prior models for their prior scores (i.e., number of edges, which were both accepted by the algorithm and included in the prior model). The models generated using the database driven priors had significantly higher accepted priors compared to the randomly generated models (p < 0.05 Student's

*t*-test for H0: μ

_{ps}

^{random}= μ

_{ps}

^{database-driven}). Note that, we performed 500 parallel BP runs using identical, database driven prior models. We observed a low degree of variation in the number of accepted priors when the models are generated using the database driven priors. This is indicative of some solution degeneracy in the models. In predictive network modeling, we accounted for the solution degeneracy by re-computing the P(W

_{ij}) at the initial step of each BP-guided decimation calculation.

**DOI:**http://dx.doi.org/10.7554/eLife.04640.018

**DOI:**http://dx.doi.org/10.7554/eLife.04640.019

**A**) The β and λ is incremented systematically while κ = λ. A low SSE with a desired connectivity level is reached for β = 2, λ = 5 and κ = 5. For different parameter sets, we obtained lower errors only for extremely connected networks inferred with high β values and so with a high weight on error in BP-optimization (e.g., β = 5, λ = 5, κ = 5, SSE = 3.8, #edges = 252). (

**B**) The β and λ is incremented while κ is fixed (κ = 5) to determine the optimum solution with better SSE and connectivity level for (κ = 5). The parameter set, β = 2 and λ = 5 yielded the most desired SSE and connectivity values. (

**C**) The κ is incremented for fixed β and λ (β = 2, λ = 5). The error is relatively insensitive to variations in κ as the final error is in the range of 5.6–6.1 for different κ values, although the connectivity increases with increasing prior strength in the cost function. This suggests that the error is mainly driven by the global β and λ terms at least around the tested β and λ values. Increasing κ, which is nonzero for only a small fraction (<2%) of possible interactions in the model, may enforce an increase in number of nonzero edges with an incremental effect on the error. κ = 5 yielded a desired connectivity value with a sufficiently low SSE. We used the parameter set (β = 2, λ = 5, κ = 5) in all BP-guided decimation calculations with the SkMel-133 data.

**DOI:**http://dx.doi.org/10.7554/eLife.04640.020

**A**) The schematic description of network simulations. The system response to paired perturbations is predicted by executing the ODE-based network models with in silico perturbations. In the ODE-based models, {W

_{ij}} represents the set of interaction strengths and is inferred with the BP-based modeling strategy. The in silico perturbations are applied as real-valued u

_{i}

^{m}vectors. The time derivative and final concentration of any predicted node is a function of the model parameters, the perturbations, and the values of all the direct and indirect upstream nodes in the models. (

**B**) The model equations are executed until all model variables (protein and phenotype responses) reach to steady state. The predicted response values are the averages of simulated values at steady state over 4000 distinct model solutions. (

**C**) The simulations expand the response map by three orders of magnitude and generate testable hypotheses. (

**D**–

**H**) The predicted phenotypic response to combinatorial in silico perturbations. Each box contains the 100 highest phenotypic responses to paired perturbations. The first box includes the response predictions for combined perturbations on primary targets (e.g., aMEK, c-Myc for G1-arrest. Also see for the definition of the term ‘primary target’). The second to sixth boxes include the predicted response for combined targeting of the primary targets with all other nodes (Nx). The last box represents the predicted response data for combination of all nodes except the primaries. For the complete predicted phenotypic response see .

**DOI:**http://dx.doi.org/10.7554/eLife.04640.005

**A**) Cell viability response to combinatorial perturbations. The desired response from a perturbation is reduction in cell viability (red: decreased cell viability, green: increased cell viability). (

**B**–

**E**) Same as in

**A**but for cell cycle progression phenotypes. For cell cycle phenotypes, the desired response is an increase in the cell cycle arrest (red: increased cell cycle arrest. green: decreased cell cycle arrest).

**DOI:**http://dx.doi.org/10.7554/eLife.04640.006

**DOI:**http://dx.doi.org/10.7554/eLife.04640.007

**A**) The isobolograms of predicted G1-response to combined targeting of c-Myc with MEK, BRAF, cyclin D1, and pJUNpS73. The leftward shift of isocurves implies synergistic interactions between the applied perturbations particularly for co-targeting of c-Myc with MEK or BRAF.

**u**denotes strength of in silico perturbations. (

**B**) RAFi inhibits MEK phosphorylation at S217 and MEKi inhibits ERK phosphorylation at T202 in a dose-dependent manner (first 2 gels). Western blot shows the level of c-Myc in response to JQ1, MEKi, RAFi, and their combinations (24 hr) (third gel). c-Myc expression is targeted with JQ1 combined with MEKi or RAFi. Direct target of JQ1, BRD4 protein is expressed in both control and 500 nM JQ1-treated cells (fourth gel) (See for uncropped western blot images). (

**C**,

**D**) The cell cycle progression phenotype in response to JQ1 and RAFi as measured using flow cytometry. 46% and 51% of cells are in G1-stage 24 hr after RAFi and JQ1 treatment, respectively. The combination has a synergistic effect on G1 cell cycle arrest (G1 = 84%). 39% of cells are in G1 when they are not treated with drugs. On panel

**D**, error bars in right panel: ±SE in three biological replicates

**E**. The drug dose–response curves of cell viability for MEKi + JQ1 (top) and RAFi + JQ1 (bottom). Cell viability is measured using the resazurin assay. Error bars: ±SEM in three biological replicates

**F**. The synergistic interactions between JQ1 and RAFi/MEKi. 1 − A

_{max}is the fraction of cells alive in response to highest drug dose normalized with respect to the nondrug treated condition (top panel). Combination index (CI) quantifies the synergistic interactions between drugs (bottom left). CI is calculated at a given level of inhibition and is a measure of the fractional shift between the combination doses (C1 and C2) and the single agent's inhibitory concentration (C

_{x,1}, C

_{x,2}).

**DOI:**http://dx.doi.org/10.7554/eLife.04640.022

**DOI:**http://dx.doi.org/10.7554/eLife.04640.023

### Publication types, MeSH terms, Substances, Grant support

#### Publication types

#### MeSH terms

- Antineoplastic Agents/pharmacology*
- Cell Line, Tumor
- Computational Biology/methods*
- Cytological Techniques/methods*
- Drug Combinations
- Drug Resistance*
- Gene Regulatory Networks
- Humans
- Melanoma/drug therapy*
- Models, Biological
- Models, Theoretical
- raf Kinases/antagonists & inhibitors