Biology.

The expression submodel depicts gene switching between an active and inactive state, transcription of the active gene to create mRNA, and translation of the mRNA to create protein. Both protein and mRNA are degraded with first-order kinetics. Although a single average gene switching rate is used for all genes [15], mRNA and protein degradation rates are informed by gene-specific half-lives taken from the literature (see S1 Table). Gene-specific transcription and translation rates are calculated from omics-data (see below). These gene-specific attributes capture dynamical behaviors that may play important roles in constraining protein function and phenotypic outcomes. Some genes possess explicitly encoded transcriptional activators and/or repressors, which can affect their transcription rate, whereas a global extrinsic effect on the translation rate of all genes is regulated by EIF4E levels (see S1 Supplemental Methods for modeling details).

Computationally efficient simulation of stochastic gene expression with expected cell-to-cell variability.

A major driver of cell-to-cell variability in phenotype and drug response, and therefore a component of intra-tumoral heterogeneity, is stochastic gene expression arising from low gene copy and mRNA number [12]. We derived a hybrid stochastic-deterministic model that captures random gene switching and mRNA synthesis/degradation, but treats as deterministic protein synthesis/degradation and the associated protein-level biochemistry since they involve much higher molecule number (Figs 2A and S2A and S2B). This model framework introduces natural noise into gene expression over time (Fig 2B). This expression submodel produces typical coefficient of variation (CV) of mRNA and protein levels [16–19] across a simulated population of cells (Fig 2C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Context tailoring and stochastic gene expression.

A) Gene copy number, mRNA levels, and protein levels inform initial conditions and expression submodel parameters. Gene switching, mRNA synthesis, and mRNA degradation are stochastic, whereas all other processes are deterministic, with a 30 second update interval. B) Example simulations for ERK. Two ERK isoform genes (MAPK1-blue and MAPK3-red; top two plots) switch between active and inactive states, leading to transcription noise (third plot down). Transcripts are summed into the signaling conglomerate ERK (fourth plot down). C) Coefficient of variation (CV) of protein (red) and mRNA (blue) levels across a population of 100 stochastically simulated cells after 24-hour serum-starvation. D) Overview of tailoring and running model. Multi-omic data from serum-starved cells sets initial conditions (“PREP”). Then, a deterministically-simulated initialization procedure maintains consistency with experimental observations in the presence of basal signaling activities resulting in a serum-starved single average cell at steady-state. From there, we run either deterministic or stochastic simulations. Prior to simulating any stimuli for stochastic simulations, we create a population of N number of cells by running the model for a 24-hour period in the serum-starved state an N number of times.

https://doi.org/10.1371/journal.pcbi.1005985.g002

Tailoring to omic-data in the MCF10A context.

Gene copy numbers in MCF10A (predominantly diploid) have been previously reported [21]. We serum- and growth factor-starved MCF10A cells for 24 hours and then quantified molecules per cell of mRNA and protein levels with mRNA-seq and proteomics, which show some agreement with prior studies [20] (S2F Fig). These measurements, with prior protein and mRNA stability estimates (see above and Methods), allow calculation of context- and gene-specific transcription and translation rates during the “PREP” phase of the modeling pipeline (Fig 2D). Many signaling activities consist of groups of genes (e.g. ERK1/2); we lump such expressed proteins into relevant “conglomerates” (Fig 2B and S1 Table).

Control of translation and ribosome dynamics.

When cell growth pathways are activated, translation rates (through release of EIF4E) and ribosome synthesis (through p70S6K) increase. We derived a rate law that incorporates the effect of free EIF4E on global translation rates, which cascade into translation rates for each gene product (Methods and S2C Fig). Ribosome number must be approximately doubled within the course of the cell cycle, which constrains the rate of ribosome synthesis, and also leads to cell growth which we model as increases in volume (S2D Fig). Simulations reproduce observed positive correlations in cell-to-cell variability for protein pairs (S2E Fig) [16].

Protein degradation.

As mentioned above, protein degradation is simulated deterministically, so fluctuations from low abundance proteins are missed, but we expect these to be a small component of overall protein level variability. There are protein-specific degradation rate constants constrained from the literature (S1 Table). Sensitivity analysis involving these parameters revealed that MEK and AKT inhibitor-induced death responses can be controlled by protein degradation rates (S6B Fig). We have not yet incorporated detailed mechanisms of protein degradation involving, for example, the ubiquitin/proteasome system. Fluctuations in the levels of proteins in these systems would contribute to extrinsic noise that is shared among proteins in a single cell, but might vary across a cell population. Incorporating such detail will be a focus of future work.

Biology.

The receptor tyrosine kinase (RTK) submodel, inspired by several prior models [22,23], depicts ligand-binding and dimerization reactions for many receptors that are frequently overexpressed, mutated or implicated in cancer, including the ErbB family (HER1-4), the hepatocyte growth factor receptor (cMet), the platelet-derived growth factor receptor (PDGFR), the fibroblast growth factor receptor (FGFR), the insulin-like growth factor receptor (IGFR), and the insulin receptor (INSR). Ligand stimulation, for most receptors, induces receptor dimerization and phosphorylation of intracellular tyrosine residues, which recruits adaptor proteins (Grb2/SOS, PI3K, and PLCγ) to mediate downstream signaling (see S1 Supplemental Methods for more details and S2 Table for rate laws and constants).

Ligand-receptor cooperativity.

Simulated receptor-ligand binding has reasonable agreement with experimentally reported cooperativity from previous literature (Figs 3A and S3A—see references in Table SI2 in S1 Supplemental Methods). Here, we simulate the case of ligand-match receptor expression, such that only the receptors that directly bind to each ligand are present and at appreciable levels (Fig 3A); we do this to ensure the system can reproduce the traditional cooperativity behavior for each receptor, unhampered by low receptor expression or expression of binding partners. When comparing the ligand-matched receptor expression to the native MCF10A expression context (S3A Fig), one discrepancy is HRG, showing no cooperativity here but was reported to have negative cooperativity [24]. HRG binds to ErbB3 and ErbB4 receptors which heterodimerize with ErbB2, all of which have context-specific expression. ErbB2 does not bind ligand and is therefore predicted to make cooperativity less negative for this class of homo and heterodimerizing receptors (including ErbB1/EGFR).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Unit testing.

A) Ligand-receptor cooperativity for RTK submodel (dots-deterministic simulations; lines- hill equation fit) based on ligand-matched receptor expression (each receptor expressed in isolation and at appreciable levels). B-D) μ-Western blot data (left column) used to train simulated (right column) ERK (B), AKT (C), and mTOR (pEIF4E-BP1) (D) signaling dynamics in the proliferation and growth submodel. Serum and growth factor starved MCF10A cells were stimulated with indicated doses of EGF, insulin, and combinations (High/Low EGF 10/0.01 nM, High/Low Insulin 1721/0.17 nM) for 5, 30, 180, or 360 minutes. E) DNA damage-induced p53 responses to single-stranded breaks (top), double-stranded breaks (middle), and both (bottom), with (red) and without (blue) simulated repair. F) Stochastic simulations for how the number of p53 pulses (top), pulse height (middle) and width (bottom) depend on DNA damage level. G) Increasing amounts of TRAIL (increasingly warmer colors) decreases time to death (as indicated by cPARP spike) exponentially (right panel: red dots-simulation; line-exponential fit). H) Survival after 5 hours TRAIL treatment. Red dots-simulation; red line-sigmoid fit; black dots-flow cytometry data. I) Effects of PUMA, NOXA, ppERK and ppAKT on time-to-death in deterministic simulations. J) Simulated cell cycle initiation by increasing cyclin D mRNA levels in serum-starved cells indicates proper oscillatory behavior of cyclin-CDK complexes.

https://doi.org/10.1371/journal.pcbi.1005985.g003

Receptor activation and trafficking dynamics.

Large ligand doses activate (i.e. receptor or adapter phosphorylation) receptor pools within ~min time scales that are coincident with receptor-specific trafficking dynamics occurring on the order of 5 minutes to 1 hour (S3B Fig) [26,27]. Modeled RTKs were consistent with experimental observations for surface vs. internalized receptors and previous models [28], such as the transient nature of EGF-induced signaling relative to other ligands (S3C and S3D Fig).

Biology.

The proliferation and growth submodel, designed as an amalgamation of several prior models [23,29,30] in addition to new mechanisms, depicts the binding of adaptor proteins to phosphorylated receptor tyrosine residues leading to the activation of the Ras/ERK and PI-3K/AKT signaling pathways. Receptor activation leads to internalization, signaling to the ERK pathway from early endosomes, and receptor degradation with some receptors being recycled back to the cell surface. The activation of the ERK and AKT pathways leads to the upregulation of cyclin D through the production of AP1 and cMYC, respectively. AKT activation can lead to mTOR activity, which exerts global control over translation efficiency via regulation of EIF4E. Negative feedback exists in both in the ERK and AKT pathways at the level of Raf and PI3K, respectively. Finally, both ERK and AKT mediate negative regulation of apoptosis signaling through BIM and BAD.

Receptor pathway preferences.

We incorporated receptor/adapter affinity and avidity patterns based on prior experimental literature and computational models (S1 Supplemental Methods). For example, ErbB3, IGF1-R and Ins-R all favor PI-3K/AKT signaling heavily as compared to Ras/ERK signaling, which is recapitulated by our model (S3C Fig). Because our model is tailored to the MCF10A context, certain ligand simulations (FGF, PDGF) do not induce appreciable activation of the ERK or AKT pathways (S3C Fig), likely due to low receptor expression. This observation is consistent with public data available in LINCS (S3D Fig) [25], with the exception of IGF, which has greater ERK activation than predicted.

Basal activity levels.

The model describes futile cycles through the Ras and PI-3K pathways (GTP bound Ras and PI(3,4,5)P), and thus accounts for the natural thresholding and negative feedback that prevents such basal activity from influencing phenotype, regulates cellular homeostasis and provides robustness to noise. This causes basal ERK and AKT kinase activity, which constrains phenotype predictions during integrated unit testing (below).

Dose and dynamic responses.

We collected dose-dependent time course data using the microwestern array (meso-scale quantitative western blot) for the ERK (pERK), AKT (pAKT), and mTOR (pEIF4E-BP1) signaling pathways in serum-starved cells treated with to EGF, insulin or the combination (Figs 3B–3D and S4). Due to the varied and dose-dependent pathway preferences of these ligands, we reasoned this dataset would provide constraints on the downstream fluxes. By varying relative fluxes through activated receptors to Ras, PI-3K, ERK, AKT and mTOR, simulations agree reasonably well with this entire cohort of data simultaneously (Fig 3B–3D). These data show the strong influence of insulin on AKT but not ERK signaling, and also transient EGF-induced signaling as reported in multiple previous studies [31].

Cell-to-cell heterogeneity in signaling response.

In order to evaluate the extent to which the model can capture cell-to-cell heterogeneity in signaling response, we compared ERK activity between simulations and experiments in response to pathway activators. We collected single-cell microscopy data of MCF10A cells stably expressing a live-cell reporter of ERK activity [32]. This kinase translocation reporter (KTR) moves from the nucleus to the cytoplasm upon phosphorylation by ERK. The cells were serum-starved then stimulated with EGF plus insulin. Qualitatively, there are two features that the model recapitulates from the imaging data (S5A and S5B Fig). First is approach to a steady-state that is slightly above baseline signaling levels. Second is the uniformity and unimodality of the responses. Other studies have reported binary on-off type switching in different cell systems [33] or in MCF10A cells but in a state of chronic growth medium treatment is asynchronous cells [34]. However, in this phase of acute response to growth factors, ERK activity seems to have uniform and unimodal response, which the model captures. Quantitatively, this imaging data is discrepant from the microwestern data. Because the microwestern assay was quantitatively validated, we do not require fold-change agreement with this imaging data.

Biology.

The DNA damage submodel, derived primarily from Batchelor et al. [35], depicts the activation of ATR and/or ATM by single/point damage and/or double strand breaks, which results in oscillatory or sustained p53 dynamics, respectively. The activation of p53 leads to MDM2 and WIP1 production, which serve as negative feedback regulators on p53 activity. Here, we model etoposide as the principal stimulus of DNA damage, which induces damage primarily when cells are cycling (see S1 Supplemental Methods). We incorporated DNA repair to either single/point damage (MSH6, MGMT) or double (BRCA2) strand breaks (Fig 3E). We required that levels of damage significant enough to induce p53 response should be repaired within the hours-day time scale.

Stochastic regulation of p53 pulsing amplitudes and dynamics.

Double strand breaks cause p53 pulses of roughly equal amplitude and timing, but increase in number as DNA damage dose increases, with this number varying across a cell population [36]. We specified ultrasensitive activation of ATM and ATR by DNA damage and tuned DNA repair activities (properties hypothesized by [36]) to ensure our model recapitulated these features (Figs 3F and S3E).

Biology.

The apoptosis submodel, derived primarily from Albeck et al. [37], depicts the activation of initiator caspases (caspase 8 and 10) by TRAIL through the death receptors (DR4/5), which leads to the activation of executioner caspases (caspases 3 and 7), leading to cell death. The activation of caspase 8 can also lead to the activation of Bax via tBid activity. Bax oligomerization drives pore formation in the mitochondria and the release of cytochrome C, which provokes apoptosome formation, the activation of caspase 3, and subsequently, the cleavage of PARP (cPARP), a common readout for apoptosis [38]. In addition to extrinsic apoptosis, there are also intrinsic (intracellular) modes of apoptosis signaling, which act through the regulation of pro- (e.g. BAD, Bim) and anti- (e.g. Bcl2) apoptotic proteins by pro-survival (active ERK and AKT) or pro-death (PUMA and NOXA, upregulated in response to DNA damage) proteins. This regulation acts through the mitochondrial-dependent pathway to promote or prevent cell death.

Extrinsic pathway dose and dynamic response.

The time-to-death of a cell decreases with increasing TRAIL dose [37], which simulations reproduce (Fig 3G), like the source model. Stochastic simulations match our percent death data for the TRAIL dose response (Fig 3H), similar to previous work [39].

Survival signaling and intrinsic pathway.

DNA damage causes p53-mediated upregulation of pro-apoptotic PUMA and NOXA, but survival signaling activity of the ERK or AKT pathways can counteract this stimulus. The model captures these general trends of PUMA/NOXA and ERK/AKT activities on time-to-death (Fig 3I).

Biology.

The cell cycle submodel, derived primarily from Gérard and Goldbeter [40], is initiated by the upregulation of cyclin D mRNA by AP1 and cMyc in G0 phase. Active Cyclin D (bound to Cdk4 or Cdk6), phosphorylates Rb to de-repress E2F transcription. E2F upregulates cyclin E and cyclin A, which bind to Cdk2 to become active, marking the beginning of S-phase and lasting into G2 phase. Finally, the activation of cyclin B/Cdk1 marks the beginning of mitosis, which, when complete, returns the cells to G1 phase.

Expected order and timing of cyclin/cdk complex activities by Cyclin D induction.

Simulated Cyclin D transcriptional activation initiates the cell cycle (Fig 3J). As in the original model, the order and timing of Cyclin expression is maintained [40], and these orders correspond to distinct phases of the cell cycle (Fig 3J). However, stochastic gene expression caused cell cycle deregulation, as found previously [41], suggesting as-yet undiscovered mechanisms of cell cycle robustness. We could not rectify this so cell cycle species do not couple to stochastic gene expression.

Adjusting translation due to protein dispersion.

After translation, proteins undergo post-translational modifications (PTM) and/or are recruited into complexes, which can change degradation rates (e.g. GSK-3β destabilizes β-Catenin). Thus, in the integrated model, total protein levels no longer agree with proteomics data. We therefore iteratively adjust effective translation rate constants to retain such agreement; this occurs during the initialization procedure (Fig 2D and S1 Supplemental Methods).

Cell fate dynamics.

Flow cytometry showed that in a serum- and growth factor-starved cell state, approximately 10% more cells die every 24 hours (Fig 4A), and active cell cycling is negligible. These data constrained the effects of proliferation and growth submodel outputs (primarily ERK and AKT activities) on cyclin D expression (cell cycle) and BAD/BIM (apoptosis), and thus the effects of basal signaling activities on phenotype.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Etoposide response dynamics in dividing and non-dividing cells.

A) Experiments (left, grey, flow cytometry) and simulations (right, blue) for MCF10A cells treated with etoposide (100μM) for 24, 48, and 72 hours in the presence of no growth factors (top) or EGF (20ng/mL) + insulin (10μg/mL) (bottom). B-C) Stochastic simulations for (B) serum-starved or (C) EGF + insulin with 72 hour etoposide treatment. D) Quantification of simulation results from (C) for wild type (blue) or p21/Chk1 loss cells (red).

https://doi.org/10.1371/journal.pcbi.1005985.g004

Data and software availability.

The MATLAB code used to run the model is available as supporting information (S1 Code).

Model implementation.

The model was implemented and simulated using MATLAB R2017b. The model switches between an ordinary differential equation portion and stochastic gene expression portion with a 30 second time step. We used a MATLAB interface to CVODES (sundialsTB v.2.4.0), which greatly increased computational speed for the deterministically simulated model portion (ODEs). Most simulations were run on a MacBook Pro with a 2.5 GHz Intel Core i5 processor and 16GB of RAM.

Initial concentrations.

In order to tailor the model to the expression context of MCF10A cells, we acquired gene copy number information from a literature source [21], mRNA expression levels by performing RNA-seq, and protein expression levels by performing mass spectrometry-based proteomics. We acquired RNA-seq and proteomics data on serum-starved cells. We used proteomics data to set initial protein levels for the RTK, proliferation and growth, and apoptosis submodels. For the majority of proteins in the cell cycle and DNA damage submodels, we used initial conditions from their original source models. One reason is that the serum-starved condition (at 24 hours) is non-cycling and low stress, with little relevant expression of such proteins. Another reason is because these submodels already included intrinsic synthesis and degradation reactions, which we largely did not alter. As exceptions to this rule, we set initial cyclin D, p21, BRCA2, MSH6, and MGMT using proteomics data, as these were either proteins we added to the submodels de novo (all except cyclin D) or served as a link between submodels (cyclin D). We acquired a value for initial ribosome level from the literature; we found a range of values from ~2 million to ~10 million copies per cell across mammalian cells [79], thus we estimated 6 million copies of ribosomes copies per cell as an intermediate value. Note this value is ~10-fold greater than the estimated number of mRNA molecules / cell, in line with the notion of the poly-ribosome (multiple ribosomes simultaneously translating the same transcript) and translation initiation being largely rate-limiting. The initial concentrations for all ligands and all post-translationally modified forms of proteins were initially set to zero. The levels of post-translationally modified forms of proteins were approximated during the initialization procedure (see Supplementary Methods for details).

Calculation of transcription and translation parameters.

Each gene exhibits a unique transcription and translation rate, which we calculated from our RNA-seq and proteomics data. The general form of the deterministic differential equations describing transcription and translation are as follows: (1) (2)

Where g* is active gene levels, m is mRNA levels, p is protein levels, k_bm is the transcription rate constant, k_bp is the translation rate constant, k_dm is the mRNA degradation rate constant, and k_dp is the protein degradation rate constant. At steady state, the transcription and translation rate constants are given by (3) (4)

The mRNA and protein degradation rate constants were calculated from gene-specific half-life measurements observed in data from [80]. Measurements from Schwanhausser et al. [80] were generated in NIH-3T3 cells, a murine fibroblast cell line; therefore, how do we know these mRNA and protein half-lives will apply to MCF10A human breast epithelial cells? First, observations point to consistent gene-specific mRNA-to-protein ratios across different human cell types [81], which, we reason, supports the conservation of protein half-lives across distinct cell types. Second, previous studies have reported conservation of mRNA and protein half-lives across species [82,80]. These reports give us confidence that mRNA and protein half-lives taken from Schwannhausser et al. [80] and other mammalian cell sources serve as reasonable estimates for half-life estimates in our MCF10A context.

We quantified the levels of mRNA and protein at steady-state (m_ss and p_ss, respectively) based on our RNA-seq and proteomics measurements. The amount of active gene product at steady-state (g*_ss) was derived. The differential equation for active gene production over time, given summation over a large population of cells, is defined as: (5)

Applying moiety conservation to gene amount yields the following steady-state condition (6)

Gene activation and inactivation rate constants were obtained from the literature [15] (see Supplementary Methods for more details). As mentioned above, the total amount of gene copies for each gene (g_total) was also obtained from a literature source [21].

Steady-states.

There are two relevant steady-states in our computations. The first is post-initialization. Here, the model is deterministic. It has expression and phenotypic components, and corresponds to a cell in a serum-starved state that is (i) not cycling, (ii) not undergoing apoptosis, and (iii) has mRNA and protein levels that match experimental measurements. With regards to cell size and ribosome number, they are constant at this steady-state. The concentrations of all molecular species in the model, in a deterministic setting, are time invariant at this steady-state as well. This corresponds to a single average cell.

The second is prior to treatment. The conditions from the above steady-state are used and then stochastic gene expression is turned on for 24 hours (see below section). This results in a cell with protein levels that have changed from the average cell and as compared to other such simulated cells.

Creating a simulated cell population in silico.

In an experimental in vitro setting, cells on a dish exhibit cell-to-cell variability in mRNA and protein expression levels. This natural variability is largely due to stochastic gene expression processes and can in some cases lead to cell-to-cell differences in phenotypic response to a perturbation. We desired to mimic this natural variability in our computational simulations. Because the model begins with the same initial conditions for each cell, we first run a stochastic simulation for each cell for 24 hours in the absence of external stimuli prior to performing any in silico experiment. This initialization step introduces a natural amount of variation in protein and mRNA levels across the simulated cell population, which randomizes the serum-starved condition in subsequent simulations (see Fig 2).

Lasso regression and support vector machine (SVM) training.

As part of our analyses we performed two lasso regressions; one related to apoptosis and one related to cell cycle. We used lasso regression to reduce our set of predictors for the SVM training. For apoptosis, we desired to identify genes that were predictive of whether a cell would die early (before 40 hours) or late (after 40 hours) in response to dual MEK and AKT inhibition. We used the lasso function in MATLAB to perform the regression. As predictor variables we used the total observable quantities of each protein (sum across all forms of a protein, i.e. conserved moiety), either at the initial time point, or averaged across a certain time frame (8 or 40 hours). Our response variable was binary (whether or not the cell died before 40 hours). We used the default settings and set the “CV” parameter (fold cross validation) to 5. Top hits were identified for mean 40 hours (Bcl2, BIM, EGFR, ErbB2, ErbB3), mean 8 hours (Bcl2, BIM, BAD, PDK1, FOXO), and initial levels (Bcl2, BIM, PARP, PDK1, FOXO). We chose to focus on time-averaged Bim and Bcl2 as they consistently appeared in all groups and no difference was found when comparing model predictability when other predictors were included or not. We then trained a binary SVM classifier using the function fitcsvm in MATLAB, using only time-averaged Bim and Bcl2 total protein levels as predictor variables and early/late death responses as the response variables. We set the “Standardize” option to “true”, which normalizes the predictor data. We used a total of 200 cells for training and another 200 for testing (for the lasso regression, above, we used the 200 training set as well). We tested the model using the predict function and generated ROC curves using the perfcurve function in MATLAB. For the cell cycle analysis, we performed the same steps as above though only considered the initial concentrations of each protein as predictors. Our response variable was binary depending on whether the cell divided within the 30-hour time course or not. CRaf and BRaf were the top two hits from the lasso regression, which we then used to train a binary SVM classifier, and generate ROC curves as above.

Sensitivity analysis on degradation parameters.

Parameter sets for degradation rate constants were generated using the lhsnorm (latin hypercube sample with normal distribution) function in MATLAB, with means equal to the original parameter and a standard deviation of 3.16 (assuming no covariance between parameters). This introduced variability of ~2 to 4-fold from the original values. The model was re-initialized for each parameter set; some sets did not meet the initialization criteria and were discarded. A random population of 25 cells was created for each parameter set (24 hours simulation) and percentage of cell death was recorded 24, 48, and 72 hours post EGF (20ng/mL) + insulin (10μg/mL) + MEK inhibitor (10μM) + AKT inhibitor (10μM) simulated treatment.

Tailoring model to U87 cell data.

To tailor the model to U87 data, mRNA molecules per cell from MCF10A cells were switched out for mRNA molecules per cell from serum-starved U87 cells (experiments and quantification performed the same as for MCF10A cells). As U87 cells are PTEN-deficient [55–57], we set its translation rate constants to zero, essentially removing it from the system. All remaining parameters were kept the same as the MCF10A-tailored and trained model. The U87-tailored model was initialized (see Fig 2D and Chapter 3 in S1 Supplemental Methods) and run in exactly the same manner as for the MCF10A-tailored model. We used gene-specific protein-to-mRNA ratios from MCF10A cells.

Significance testing.

Significance testing was performed by using the ttest2 function in Matlab. Three stars indicate a p-value less than 0.001, two stars indicate a p-value in-between 0.001 and 0.01, and one star indicates a p-value between 0.01 and 0.05. Error bars correspond to standard error of the mean where possible.

Flow cytometry software.

Analysis of flow cytometry data was performed using Cytobank. Raw flow cytometry data is available at flowrepository.org, accessions FR-FCM-ZYGG, FR-FCM-ZYHQ, FR-FCM-ZYHP, FR-FCM-ZYHX, and FR-FCM-ZYGF.

Cell culture.

MCF10A cells were cultured in DMEM/F12 (Gibco; Cat: 11330032) medium supplemented with 5% (v/v) horse serum (Gibco; Cat: 16050–122), 20ng/mL EGF (PeproTech, Cat: AF-100-15), 0.5 mg/mL hydrocortisone (Sigma, Cat: H-0888), 10μg/mL insulin (Sigma, Cat: I-1882), 100ng/mL cholera toxin (Sigma, Cat: C-8052), 2mM L-Glutamine (Corning; Cat: 25-005-CI) and 1X (10,000 I.U./mL / 10,000 ug/mL) penicillin/streptomycin (Corning; Cat: 30-002-CI). Cells were cultured at 37°C in 5% CO₂ in a humidified incubator, and passaged every 2–3 days with 0.25% trypsin (Corning; Cat: 25-053-CI) to maintain subconfluency. Cells originate from a female host, were a gift (Gordon Mills and Yoko Irie) and we did not independently authenticate the cells. Starvation medium was DMEM/F12 medium supplemented with 2mM L-Glutamine and 1X penicillin/streptomycin.

RNA sequencing and analysis.

MCF10A cells were seeded at ~30% density (1 million cells / 60mm diameter dish, one dish per biological replicate), and incubated overnight in full growth medium. The next morning, the full growth medium was aspirated, cells were washed once with PBS, and then placed in starvation medium for 24 hours. Total RNA was extracted using TRIzol (Life Technologies, Cat: 15596018) per manufacturer instructions (detailed SOP at www.dtoxs.org; DToxS SOP A– 1.0: Total RNA Isolation). RNA sequencing and analysis was performed as previously described [83]. We prepared RNA-seq libraries by adapting a single-cell RNA-seq technique—Single Cell RNA Barcoding and Sequencing method (SCRB-seq; [84,85])—to extract total RNA from cells (detailed SOP at www.dtoxs.org; DToxS SOP A– 6.0: High-throughput mRNA Seq Library Construction for 3' Digital Gene Expression (DGE)). This method involves converting each Poly(A)+ mRNA molecule to cDNA labeled with universal adaptors, sample-specific barcodes, and unique molecular identifiers (UMIs) on every transcript molecule using a template-switching reverse transcriptase. cDNA was amplified and prepared for multiplexed sequencing, enriching for 3’ ends preserving strand information by using a modified transposon-based fragmentation approach. Libraries were sequenced using the Illumina HiSeq 2500 platform (detailed SOP at www.dtoxs.org; DToxS SOP A– 7.0: Sequencing 3'-end Broad-Prepared mRNA Libraries). Analysis was performed using a custom Python script (detailed SOP at www.dtoxs.org; DToxS SOP CO– 3.1: Generation of Transcript Read Counts). In brief, sequenced reads were aligned to a 3’-focused hg19 reference using BWA, and counts of reads that aligned to one or more than one gene with high confidence were calculated. The UMI count is defined as the number of distinct unique molecular identifiers (UMI) sequences embedded in those aligned reads and is well-known to alleviate PCR bias that can distort linearity in quantification. Biological triplicates were performed, but one sample failed to be sequenced, leaving biological duplicates.

Proteomics and analysis.

MCF10A cells were seeded at ~30% density (5 million cells / 150mm diameter dish, 2 dishes pooled per biological replicate) and incubated overnight in full growth medium. The next morning, the full growth media was aspirated, cells were washed once with PBS, and then placed in starvation medium for 24 hours. Cells were trypsinized for 10 minutes, resuspended in 1X PBS, and spun down at 500g for 5 minutes. Cell pellets were then snap frozen in liquid nitrogen and shipped to the Advanced Proteomics Center at New Jersey School of Medicine to perform mass spectrometry. The cell pellet was subjected to urea lysis buffer composed of 8 M urea, 100 mM TEAB, 1X protease inhibitors (complete, EDTA-free Protease Inhibitor Cocktail from Sigma, prepared according to manufacturer instructions) and 1X phosphatase inhibitors (PhosSTOP from Sigma, prepared according to manufacturer instructions). Protein (50.0 μg total) was resolved by one dimensional SDS-PAGE. The gel lane was divided into 24 equal size bands, followed by in-gel digestion with trypsin performed as in [86]. Peptides were fractionated by reverse phase chromatography and analyzed directly by LC/MS/MS on a Q Exactive Orbitrap mass spectrometer (funded in part by NIH grant NS046593, for the support of the Rutgers Neuroproteomics Core Facility). We used Maxquant to analyze the data and calculate iBAQ scores, using a human UniProt protein database as the reference proteome. Final proteins were identified with 1.0% false discovery rate (FDR) at both protein and peptide levels. This resulted in ~10k unique proteins. Biological triplicates were performed.

Calculation of mRNA and protein molecules per cell.

Importantly, the units of the data acquired from the RNA-seq (UMI count; [84]) and mass spectrometry (iBAQ scores; [87]) technologies are linearly proportional to the number of mRNA or protein molecules in the original sample, respectively. To convert from these measurement units to molecules per cell, we first quantile normalized the three biological replicates using the MATLAB function, quantilenorm, with default settings. The quantile-normalized values were then converted from their original unit of measure to units of molecules per cell with the following logic.

Consider measurement of mRNA or protein i, A_i (in UMI counts or iBAQ score), and the corresponding absolute copy number per cell that gives rise to this measurement, X_i. Due to the linearity of the measurement modality, we have: (7)

Summing over all species n in the cell gives: (8)

Dividing by these sums then yields the following proportionality to convert the measurement to absolute copy number per cell.

(9)

Using these equations requires an estimate for the total number of mRNA and protein molecules in a cell. The total number of mRNAs in a typical mammalian cell was estimated from three sources. The first was taken from [80] by summing all mRNA quantities found in their search (173,921 molecules per cell). The second was taken from the Molecular Biology of the Cell textbook [88], which gave a value of 360,000 molecules/cell. The third was calculated by converting the total picogram amount of RNA in a typical mammalian cell, estimated from one source to be 26pg [89], to molecules per cell given estimates that ~3% of total RNA is mRNA [79], that there are an average of 2000 nucleotides in a single mRNA transcript [79], and that the average molecular weight of a single nucleotide is 339.5 g/mol [79]; using this information we calculated a total 691,891 molecules per cell. We took the average of these three numbers and approximated the result to 400,000 molecules per cell. To approximate the total number of proteins in a typical mammalian cell we summed the total number of protein molecules found by Schwanhausser et al. [80], which amounted to 3.08x10⁹ protein molecules per cell. Importantly, our method for converting to molecules per cell resulted in reasonable agreement with data from Schwanhausser et al. [80]. When comparing the numbers between our dataset (“ours”) and the Schwanhausser dataset (“S”) for a handful of constitutively expressed proteins that typically do not significantly change across cell types, we find most of them to be within the same order of magnitude: MAP2K1 (ours: 147250; S: 284958), MAP2K2 (ours: 377540; S: 327073), MAPK1 (ours: 746463; S: 1400595), MAPK3 (ours: 76314; S: 195819), AKT1 (ours: 74762; S: 37060), and AKT2 (ours: 81476; S: 4325). Discrepancies here could be due to cell-type differences, however.

Other studies have integrated universal spike-in standards (e.g. [20]), which allow for a more direct experimental determination of absolute protein copy numbers per cell. Here we did not integrate such standards but rather used the method described above based on label-free quantification by iBAQ score. A subset of the published spike-in data was directly comparable with our data. The comparison demonstrates reasonable agreement (S2F Fig).

Micro-western experiments and analysis.

MCF10A cells were seeded at ~30% confluence in 6-well plates (150,000 cells / well), incubated overnight in full growth medium. The next morning, the full growth medium was aspirated, cells were washed once with PBS, and then placed in starvation medium for 24 hours. Cells were treated with combinations of EGF and insulin at various doses for 5 minutes, 30 minutes, 3 hours, or 6 hours. Growth medium was aspirated fully and cells were lysed on ice with a custom-made lysis buffer tailored for micro-western applications composed of 240 mM Tris-Acetate, 1% SDS w/v, 0.5% glycerol v/v, 5mM EDTA, water, and various protease and phosphatase inhibitors—aprotinin (1μg/mL), leupeptin (1μg/mL), pepstatin (1μg/mL), β-glycerophosphate (10mM), sodium orthovanadate (1mM), and dithiothreitol (DTT; 50mM). Reagent sources are listed in our SOPs (www.birtwistlelab.com/protocols). Samples were boiled at 95 degrees for 5 minutes, sonicated using a VialTweeter (Hielscher), then concentrated ~5 to 10-fold using Amicon Ultra centrifugal filters (Merck Millipore Ltd). Total protein content was assayed via the Pierce 660 assay per the manufacturer’s instructions. Samples were printed in 24 “blots” onto a single 9.5% poly-acrylamide gel using a GeSim Piezoelectric Nanoplotter, followed by horizontal electrophoresis and wet transfer to a nitrocellulose membrane. The membrane was blocked with Odyssey Blocking Buffer (TBS; LICOR) then put into a 24-well gasket system that corresponds to the original lysate printing and horizontal running pattern. This gasket system allows different primary antibodies to be probed in each well. We probed using the following antibodies from Cell Signaling: pERK (Cat: 4370; 1:1000), pAKT (Cat: 4060; 1:2000), pEIF4E-BP1 (Cat: 2855; 1:500), cyclin D (Cat: 2978; 1:500) and α-Tubulin (Cat: 3873; 1:2000). We then incubated the entire gel with two secondary antibodies from LiCor: IR dye 800CW goat anti-rabbit (green; Cat: 926–32211) and IR dye 680CD goat anti-mouse (red; Cat: 926–68070) at 1:20,000. The membrane was then imaged via a LICOR Odyssey scanner and quantified with Image Studio Lite software as “signal” (intensity with background subtracted). For full SOPs, please visit www.birtwistlelab.com/protocols. Each signal (pERK, pAKT, pEIF4E-BP1, and cyclin D) was first divided by the respective α-Tubulin signal (loading control) in the same lane to normalize for lane-to-lane differences in loading. Then to normalize for any systematic inter-well differences, each ratio was then normalized to the serum-starved control lane in its well. Raw images can be found in S4 Fig.

BrdU incorporation/propidium iodide flow cytometry experiments and analysis.

MCF10A cells were seeded at ~1 million cells in a 10cm dish, and incubated overnight in full growth medium. The next morning, the full growth media was aspirated, cells were washed once with PBS, and then placed in starvation medium for 24 hours. Cells were treated with indicated doses of EGF (10nM), insulin (10μg/mL), the combination dose (10nM EGF + 10μg/mL insulin), and 30 min pre-incubation with a MEK inhibitor (PD0325901, Sigma, Cat: PZ0162) or AKT inhibitor (MK2206, ChemieTek, Cat: CT-MK2206) as indicated (both at 10μM). An hour before lifting, cells were treated with a pulse of 10 μM BrdU (Sigma). Cells were trypsinized, washed with 4°C 1X PBS and pelleted by centrifugation at 1000rpm for 5 minutes. To fix cells, they were resuspended in 200uL 4°C 1X PBS and 2mL of 4°C 70% ethanol (v/v in ddH₂0) were added dropwise while vortexing at low speed. The cells were then stored at 4°C for at least 30 minutes. Cells were pelleted at 1000rpm for 5 minutes and resuspended in 1mL PBS. We added 1mL of 4N HCl and incubated the cells at room temperature for 15 minutes. We washed the cells by centrifuging at 1000rpm for 5 minutes and resuspending in 1mL PBS. After another spin, we resuspended the cells in 1mL PBT (PBS + 0.5% BSA (g/L) + 0.1% Tween 20 (v/v) (Sigma)). We spun the cells down and resuspended in 200 μL of PBT containing 1/40 dilution of anti-BrdU antibody (BD Biosciences) and incubated at room temperature for 30 minutes. We then added 1mL of PBT, spun the cells down and resuspended in 200uL PBT containing 1/40 dilution of anti-mouse FITC conjugate (BD Biosciences); we incubated cells in this for 30 minutes at room temperature in the dark. We added 1mL of PBT, spun cells down and resuspended in 250 μL PBS containing 250μg/mL RNAseA (Qiagen MaxiPrep Kit) and 10μg/mL propidium iodide (Sigma). We incubated the cells at room temperature for 30 minutes in the dark prior to analysis by flow cytometry on an Accuri C6 (BD Biosciences). We first gated for non-debris cells (FSC-A vs. SSC-A), then gated for singlets (SSC-A vs. SSC-H), then plotted PI intensity (x-axis) versus BrdU intensity (y-axis). We quantified percentages of cells in G1/G0 phase (First PI peak, BrdU negative), S-phase (BrdU positive), and G2/M phase (2^nd PI peak, BrdU negative). Our metric of percent cell cycling was the S-phase population plus the G2/M population percentages. Raw data are available at flowrepository.org.

Annexin V/Propidium iodide flow cytometry experiments and analysis.

MCF10A cells were seeded at ~30% confluence in 12-well plates (75,000 cells / well), and incubated overnight in full growth medium. The full growth medium was replaced by starvation medium for 24 hours. Cells were then treated with different growth factor-inhibitor combinations as indicated and cells were assayed at 24, 48, and 72 hours post stimulation. A MEK inhibitor (PD0325901, Sigma, Cat: PZ0162) and/or AKT inhibitor (MK2206, ChemieTek, Cat: CT-MK2206) was used at 10μM. Anisomycin (Sigma, Cat: A9789), an inhibitor of protein and DNA synthesis, used at 10μM served as a positive control for cell death (death percentages were always greater than 90% cell death). Cells in full growth medium without treatment served as a negative control (death percentages were always less than 10% cell death). Cells were trypsinized, resuspended in PBS and pelleted by centrifugation at 1100g (to ensure collection of dead cells) for 5 minutes at room temperature. Cells were then resuspended in 100uL of Annexin-V binding buffer (10mM Hepes, 140mM NaCl, 2.5mM CaCl₂ (Sigma) in ddH₂0 pH 7.4) with 5uL of AnnexinV-FITC (BioLegend, Cat: 640906) and 0.1μg/mL propidium iodide (Sigma). Cells were incubated for 20 minutes at room temperature prior to analysis by flow cytometry. In TRAIL experiments, cells in full growth medium were treated with multiple doses of superkiller TRAIL (tumor necrosis factor (TNF)-related apoptosis-inducing ligand) (Enzo Life Sciences, Cat: ALZ-201-115-C010). After 5-hour TRAIL treatment, cells were harvested and analyzed by flow cytometry as described above, and a dose-response curve of two-fold serial dilutions of TRAIL ranging 2-300ng/mL was plotted. We first gated for non-debris cells, then gated for singlets (both as above), then quantified the % of cells in the Annexin-V positive population, which was our metric for percent cell death. Raw data are available at flowrepository.org.

Etoposide flow cytometry experiments.

MCF10A cells were seeded in full growth medium at approximately 30% confluence in 12-well plates (75,000 cells / well) and incubated overnight. Cells in full growth medium were then treated with 100 μM of Etoposide (Cayman Chemical, Cat: 12092) and assayed at 24, 48, and 72 hours post stimulation. Alternatively, after the cells attached overnight, the full growth medium was aspirated, cells were washed once with PBS, and then placed in starvation medium for 24 hours. These cells were then treated with 100 μM Etoposide and assayed at 24 and 48 hours post stimulation. Again, Anisomycin (10μM; Sigma, Cat: A9789) was used as a positive control for cell death and cells in full growth medium without treatment were used as a negative control for cell death. Cells were assayed with Annexin V-FITC and PI and analyzed as described above for the apoptosis flow cytometry experiments. Raw data are available at flowrepository.org.

Single-cell microscopy experiments.

Kinase translocation reporters (KTR) are fluorescent protein-linked kinase translocation biosensors that shuttle between the nucleus and cytoplasm based on the activation state of the pathway of interest. For this ERK KTR construct, the probe is localized to the nucleus when inactive (via NLS) and shuttled to the cytoplasm when active/phosphorylated by ERK (via NES). Signaling dynamics can be determined by plotting the ratio of cytoplasmic-to-nuclear KTR fluorescence [32].

The ERK KTR clover probe (pLenti CMV HYGRO DEST ERK KTR Clover; hygromycin resistance) was generated using gateway cloning enzymes BP, LR clonase, and the source vector pLenti CMV Puro DEST ERK KTR Clover. Lentivirus was generated as previously described [90] for both this ERK KTR clover probe and a pNLS-iRFP713 construct (used as nuclear marker). Lentiviral supernatant was collected and concentrated via Amicon Ultra-15 filters NMW 100 kD centrifugal filters.

MCF10A cells were transduced with ERK KTR concentrated lentiviral supernatant. Twenty-four hours post transduction, cells were imaged to confirm expression of ERK KTR clover. Cells expressing ERK KTR clover were sparsely seeded on a 150x25mm dish and cultured in selection media, DMEM F12, 5% Horse serum, EGF (20ng/mL), Hydrocortisone (.5μg/mL), Cholera toxin (100ng/mL), Insulin (10μg/mL), and hygromycin (31.25μg/mL) and placed in a humidified incubator at 37°C, 5% CO2. Cells were cultured until colonies were visible under brightfield microscope. Individual colonies were harvested by trypsonization (.25%), re-plated and imaged to verify expression of ERK KTR clover. Colonies showing uniform expression of ERK KTR were used in subsequent probe validation.

MCF10A cells were seeded at ~30% confluence (5,000 cells/well) in 96-well plates (Corning 3603) and incubated overnight in full growth medium (DMEM F12, 5% Horse serum, EGF (20ng/mL), Hydrocortisone (.5μg/mL), Cholera toxin (100ng/mL), Insulin (10μg/mL), and hygromycin (1.5μg/mL)). These cells were then transduced with the pNLS-iRFP713a nuclear marker (used for image segmentation purposes). The following day, the full growth medium was replaced by starvation medium for 24 hours. We then treated cells with EGF (20ng/mL) plus insulin (10μg/mL) and quantified ERK KTR clover response over time at 3 minute intervals over the course of 1 hour. We compared this to mock treatment (no treatment; serum-starved) and cells pre-treated with a MEK inhibitor (PD0325901; 100nM) for 30 minutes following stimulation with EGF plus insulin. Baseline ERK KTR activity was acquired for each condition for 15 minutes. Images were acquired with the InCell2200 (incubated at 37°C with CO2, Nikon 20X/0.75, Plan Apo, CFI/60, 3 fields/condition, KTR Ex. 475/28 Em. 511.5/23, iRFP Ex. 632/22, Em. 684/25, brightfield). Images were analyzed using a custom CellProfiler pipeline [91], segmenting cytoplasmic and nuclear compartments for each cell. Baseline normalized cytoplasmic-to-nuclear KTR fluorescence was plotted for each segmented cell. The CellProfiler pipeline can be supplied when requested.

Additional resources.

A full SOP for the μ-Western blot procedures can be found at www.birtwistlelab.com/protocols. More detailed SOPs for RNA isolation (DToxS SOP A– 1.0: Total RNA Isolation), RNA-seq library preparation (DToxS SOP A– 6.0: High-throughput mRNA Seq Library Construction for 3' Digital Gene Expression (DGE)), RNA sequencing (DToxS SOP A– 7.0: Sequencing 3'-end Broad-Prepared mRNA Libraries), and analysis (DToxS SOP CO– 3.1: Generation of Transcript Read Counts) can be found at www.dtoxs.org.