A mechanistic model captures the emergence and implications of non-genetic heterogeneity and reversible drug resistance in ER+ breast cancer cells

Abstract

Resistance to anti-estrogen therapy is an unsolved clinical challenge in successfully treating ER+ breast cancer patients. Recent studies have demonstrated the role of non-genetic (i.e. phenotypic) adaptations in tolerating drug treatments; however, the mechanisms and dynamics of such non-genetic adaptation remain elusive. Here, we investigate coupled dynamics of epithelial-mesenchymal transition (EMT) in breast cancer cells and emergence of reversible drug resistance. Our mechanism-based model for underlying regulatory network reveals that these two axes can drive one another, thus enabling non-genetic heterogeneity in a cell population by allowing for six co-existing phenotypes: epithelial-sensitive, mesenchymal-resistant, hybrid E/M-sensitive, hybrid E/M-resistant, mesenchymal-sensitive and epithelial-resistant, with the first two ones being most dominant. Next, in a population dynamics framework, we exemplify the implications of phenotypic plasticity (both drug-induced and intrinsic stochastic switching) and/or non-genetic heterogeneity in promoting population survival in a mixture of sensitive and resistant cells, even in the absence of any cell-cell cooperation. Finally, we propose the potential therapeutic use of mesenchymal-epithelial transition inducers besides canonical anti-estrogen therapy to limit the emergence of reversible drug resistance. Our results offer mechanistic insights into empirical observations on EMT and drug resistance and illustrate how such dynamical insights can be exploited for better therapeutic designs.

Figure 1.

Emergent dynamics of coupled EMT-ERα signaling network. (A) Gene regulatory network (GRN) showing crosstalk between EMT and Estrogen receptor (ERα) signaling. Green arrows represent activation links; red hammers represent inhibition. (B) Heatmap of stable steady-state solutions for network shown in (A), obtained via RACIPE. (C) Density histogram of EM Score ( = Zeb1 – miR-200) and Resistance Score ( = ERα36 – ERα66) fitted to 3 and 2 Gaussian distributions, respectively. Red dotted lines show segregation between phenotypes: Epithelial (E), Hybrid (H) and Mesenchymal (M) for EM score, and Resistant (R) and Sensitive (S) phenotype for the latter. (D) UMAP dimensionality reduction plots for steady-state solutions states obtained by RACIPE colored by either EM Score or Resistance score. (E) Scatter plot showing corresponding EM Score and resistance score for all RACIPE solutions, and six biological phenotypes. Spearman correlation coefficient (ρ) and corresponding P-value are reported. (F) Gene expression levels in six biologically defined phenotypes. Dotted line represents the monotonic increase in levels of ZEB1 across the phenotypes. Standard deviation is plotted as error bars. (G) Density distribution of Spearman correlation coefficients (ρ) across an ensemble of 100 randomized versions of the GRN shown in (A). Red line shows the correlation coefficient of the wild-type network shown in (A).

Figure 2.

Gene expression analysis of publicly available datasets. (A) Correlation of ESR1 with ZEB1, SNAI2 (SLUG) and activity of MSigDB Hallmark EMT signature (ssGSEA scores) in a cohort of 87 ER+ breast cancer patients (GSE6532). (B) Correlation of tamoxifen resistance (ssGSEA scores) signature with expression of ZEB1 and SNAI2 and activity of hallmark EMT signature in primary breast tumors treated with tamoxifen in adjuvant setting (GSE9195). (C) Correlation of ESR1 expression levels and estrogen response activity with CDH1 and VIM in tumor samples from 298 ER+ patients treated with tamoxifen for 5 years. (D) Diagonal correlation matrix between expression levels (ZEB1, SLUG, VIM, CDH1, ESR1), EMT scoring metrics (76GS, MLR and KS) and gene set activity estimation (ER early response, ER late response, tamoxifen resistance, hallmark EMT signatures) for 60 samples of micro-dissected tumour biopsies (GSE1378) and whole tissue tumour biopsies (GSE1379) from a cohort of patients treated with tamoxifen for 5 years. (E) Correlation plots of estimated activities of estrogen response with hallmark EMT signatures in different subtypes of breast cancer in TCGA. Spearman correlation coefficient (ρ) and corresponding P-value are reported.

Figure 3.

Induction of EMT can drive suppression of estrogen signaling and vice versa. (A) Impact of over-expression/down-expression of ZEB1 levels in RACIPE simulations on frequencies of different biological phenotypes. Error bars denote standard deviation across n = 3 replicates. (B) Experimental data (GSE43495) for EMT induction via Twist, Snail or Slug in HMLE cells and the concurrent decrease in the magnitude of early and late estrogen response. (C) Experimental data showing EMT induction via TGFβ, Twist, Gsc and Snail in HMLE epithelial cells and the concurrent decrease in the magnitude of early and late estrogen response (GSE24202). (D) Same as (A) but for over-expression/down-expression of ERα66. (E) Experimental data showing differences in gene expression levels of Cdh1, Vim, Snai2 (Slug), Vim and Zeb1 and change in magnitude of early and late estrogen response and the EMT program (ssGSEA on MSigDB hallmark EMT signature) in control and ERα silenced MCF7 cells (GSE27473). (F) Experimental data showing differences in activity levels of early ER response, late ER response and EMT program in sensitive and resistant MCF7 cell lines (GSE67916). For A–F, * denotes a statistically significant difference between the control and perturbed/induced case assessed by a two-tailed Student’s t-test assuming unequal variances. (G) Scatter plot showing association between activity of early estrogen response and cells with varying positions on a 2D epithelial–mesenchymal plane (GSE147356). Spearman’s correlation coefficient between epithelial and mesenchymal scores, and corresponding P-value are reported. Color bar represents the activity of early estrogen response. (H) Scatter plot showing activity of EMT signature in TGFβ treated MCF7 individual cells in pseudo time and the concurrent decrease in BRCA ESR1 regulon activity. Color bar represents the range of activity level of the ESR1 regulon (GSE147405). (I) Schematic showing bidirectional associations between the EMP and the drug resistance program, i.e. induction of EMT drives a switch to a therapy-resistant state, and acquisition of therapy resistance often drives EMT.

Figure 4.

Stochastic stimulations showing dynamic state transitions among different biological phenotypes. (A) Fraction of RACIPE parameter sets resulting in monostable and multistable solutions (bi-, tri-, others) and the frequency distribution of phases that compose the monostable and multistable solution sets. (B) System dynamics for a representative {ES, HR, MR} parameter set showing the existence of the three biological EM phenotypes (E, H, M) and resistant (R) and sensitive (S) phenotypes when started from multiple initial conditions. (C) Time course showing the transition of the system from a MR to an ES phenotype through a HR state under the influence of noise. Sensitivity score is defined as negative of the tamoxifen resistance score, i.e. ERα66–ERα36. (D) Marginal distribution of the EM score from the time course shown in (C); three peaks denote existence of three distinct states along EM spectrum. (E) (top) Stochastic time series for multiple initial conditions tracking EM and Resistance scores in a representative parameter set from the {ES, HS, MR} phase. (Bottom) Landscape obtained by simulation of that parameter set with valleys representing stable states possible in the system. (F) Same as (E) but for a representative parameter set from the {ES, HR, MR} phase. (G) Changes in SLUG levels as the system transitions from ES to MR phenotype through either HS or HR state. HR state is characterized by high levels of SLUG compared to HS cell state. Student’s t-test results show the level of statistical significance between various comparisons.

Figure 5.

Effect of heterogeneity and plasticity on tumor survival in the presence of an anti-estrogen drug. (A) Schematic for model formulation showing inter-conversions between sensitive and resistant phenotypes (transition probabilities: P_SR, P_RS). Heterogeneity in the cell population is modelled by standard deviation (SD) of Gaussians from which resistance scores are sampled. Survival probability of cells is a function of resistance score approximated as a sigmoidal curve (shown in red). (B) Effect of heterogeneity on population sizes over time, starting with an initially all sensitive cell population and at P_SR = P_RS = 0. (C) Effect of plasticity on population sizes over time starting with an initially all sensitive cell population and no heterogeneity (SD = 0). (D) Population sizes (at time [t] = 100) as a function of P_SR and P_RS at two different heterogeneity levels, starting with an initially all sensitive cell population. Dotted lines indicate a qualitative boundary between tumor survival and elimination scenarios. (E) Distinct qualitative scenarios—collapse of initial population of cells, maintenance of the cell population around starting initial conditions (in the time frame considered) and net growth in a population of cells leading to survival of the tumor—at varying levels of P_SR and P_RS. All simulations start with a fixed heterogeneity (SD = 0.5) and a fully sensitive population. (F) Population sizes over time as a function of varying values of (P_SR, P_RS). All simulations start with no heterogeneity (SD = 0) and a fully sensitive population. Shaded area represents the standard deviation around the mean of n=10 replicates.

Figure 6.

Different mechanisms promoting survival of a cancer cell population in the presence of an anti-estrogen drug. (A) Simulations showing increase in survival probability (decrease in extinction probability) of a cancer cell population with varying heterogeneity levels (SD = 0.65, 0.70 and 0.75) starting form an all sensitive population and fixed values of P_SR and P_RS at 0.5 and 1.0, respectively. (B) Simulations showing a modest increase in the survival probability of a cell population with varying levels of initially resistant cells (initial fraction = 0.0, 0.25 and 0.50) and fixed values of P_SR and P_RS at 0.5 and 1.0, respectively. Orange ribbon represents the collection of all states that survive and form a colony and blue ribbon for all those cases that are eventually eliminated. The dark line represents the mean and the band represents the SD around that mean for an ensemble of simulations. (C) Final population sizes (at time [t] = 100) as a function of P_SR and P_RS at two different heterogeneity levels (SD = 0, 1) and two different levels of drug induced plasticity (0 and 1) starting with an initially all sensitive cell population. Dotted lines indicate a qualitative boundary between cases where a tumour survives or is eliminated by the presence of the drug.

Figure 7.

MET inducer, in conjunction with anti-estrogen drugs, can potentially limit the survival of cancer cell population. (A) Temporal evolution of a population of cancer cells (resistant and sensitive both) under different levels of drug induced plasticity (0 and 1) and MET induced sensitivity (0 and 1). Simulations were performed starting form an all sensitive population and fixed values of P_SR and P_RS at 0.2 and 0, respectively. (B) Final population sizes (at time [t] = 100) as a function of P_SR and P_RS two different levels of drug induced plasticity (0.1 and 1) and two different levels of MET induced drug sensitivity (0 and 1) starting with an initially all sensitive cell population with no heterogeneity in the system. Dotted lines indicate a qualitative boundary between cases where a tumour survives or is eliminated by the presence of the drug.

Figure 8.

Schematic depicting dynamical traits of coupled EMT-ERα signaling network and its implications in tumor survival. (A) Landscape showing multiple phenotypes defined on EMT and drug resistance axes: ES (epithelial-sensitive) and MR (mesenchymal-resistant) phenotypes are more dominant (witnessed by depth of the valley in the landscape). Arrows induced transitions under the influence of noise or drug among six phenotypes (ES, ER, HS, HR, MS, MR). (B) Population dynamics showing multiple parallel paths to long-term resistance (pre-existing reversibly resistant cells, stochastic switching and dynamic equilibrium, and drug-induced plasticity) and the predicted effect of combinatorial therapy (anti-estrogen therapy + MET inducers) to drive population collapse.