Addressing genetic tumor heterogeneity through computationally predictive combination therapy

Abstract

Recent tumor sequencing data suggest an urgent need to develop a methodology to directly address intratumoral heterogeneity in the design of anticancer treatment regimens. We use RNA interference to model heterogeneous tumors, and demonstrate successful validation of computational predictions for how optimized drug combinations can yield superior effects on these tumors both in vitro and in vivo. Importantly, we discover here that for many such tumors knowledge of the predominant subpopulation is insufficient for determining the best drug combination. Surprisingly, in some cases, the optimal drug combination does not include drugs that would treat any particular subpopulation most effectively, challenging straightforward intuition. We confirm examples of such a case with survival studies in a murine preclinical lymphoma model. Altogether, our approach provides new insights about design principles for combination therapy in the context of intratumoral diversity, data that should inform the development of drug regimens superior for complex tumors.

Significance: This study provides the first example of how combination drug regimens, using existing chemotherapies, can be rationally designed to maximize tumor cell death, while minimizing the outgrowth of clonal subpopulations.

Fig. 1. A strategy for modeling intratumoral heterogeneity and mathematical optimization

(A) A simplified schematic of tumor evolution, with deregulated Myc expression and p19^Arf loss, followed by additional loss-of-function mutations. (B) RNAi-based modeling of intratumoral heterogeneity. Each shRNA knockdown models a specific loss-of-function event. Mixture of these subpopulations creates a heterogeneous tumor population. (C) A schematic of how mathematical optimization can be applied to drug design for tumor heterogeneity. We have previously acquired a dataset on the response of specific shRNA knockdowns to a set of single chemotherapeutic and targeted agents (23,24). Here, using this dataset and given a particular population composition, we applied an optimization approach (see Material and Methods) to determine drug combinations that are best and worst at treating all subpopulations. (D) Two top “hits” (i.e. population compositions) derived from computational simulation demonstrating that the optimal drug combination predicted is different depending on whether we examine the entire heterogeneous tumor population or only a particular subpopulation (see Fig. S2 and Table S1 for summary of all simulation results).

Fig. 2. In vitro validation of predicted effects of combination therapies on subpopulation composition

(A) A schematic of the in vitro competition assay (see also Fig. S3). Eμ-Myc; p19^Arf-/- lymphoma cells were retrovirally transduced with the desired shRNA(s). Mixed populations were treated with combination treatment and analyzed at 48h. (B) A representative flow cytometry result of no treatment and the indicated combination treatments of a mixed population of tumor cells containing parental, shChk2 (Tomato-labeled), and shBok (GFP-labeled). (C-E) In vitro competition assay results for different population composition using different combination treatments. Drug combinations predictably enriched/depleted subpopulations. (F) Correlation between predictions made from mathematical model and the actual experimental results. Gray data points represent individual subpopulations (e.g. shChk2 only) and blue data points represent combined subpopulations (e.g. shChk2 and shBok). Data shown are mean ± s.e.m. of three independent experiments, **P<0.01, ***P<0.001 (ANOVA with Bonferroni post-hoc test).

Fig. 3. In vivo validation of effects of combination therapies on subpopulation composition

(A) A schematic of the in vivo competition assay. Eμ-Myc; p19^Arf-/- lymphoma cells were retrovirally transduced with desired shRNA, and mixed populations of tumor cells were tail-vein injected into recipient mice. Combination treatments were given at presentation of palpable tumor. Tumor cells were collected and analyzed at relapse. (B) Representative fluorescence imaging of vehicle-treated mice with mixed populations of parental, shChk2 (GFP-labeled), and shBok (Tomato-labeled) infected tumor cells, showing intratumoral heterogeneity in vivo. (C) Representative in vivo competition assay flow cytometry analysis of relapsed tumors following in vivo treatment with the indicated combination therapies. (D) An in vivo competition assay showing the enrichment or depletion of subpopulations of shRNA infected tumor cells. Vin/SAHA and IRT/CBL treatment data for the combined shBok and shChk2 populations are shown in lymph node, thymus, and spleen. Drug combination predictably enriched/depleted subpopulations, in agreement with in vitro results (Fig. 2). Data shown are mean ± s.e.m of three independent experiments (with 4-5 mice per experiment). **P<0.01 (two-tailed Student's t-test). (E) Relative tumor-free survival for each combination treatment on mice transplanted with heterogeneous parental/shChk2/shBok tumor, with days normalized to the median tumor-free survival of treated mice with homogeneous parental tumor. Vin/SAHA improved relative tumor-free survival when compared to IRT/CBL. Data were compiled from four independent experiments. P value was calculated using a log-rank test.

Fig. 4. Vin/SAHA is superior to IRT/CBL in extending relative tumor-free survival in a three-component heterogeneous tumor across varying subpopulation proportions

(A) A ternary plot showing the comparison between Vin/SAHA and IRT/CBL in terms of relative tumor-free survival in a shChk2/shBok/parental tumor (i.e. tumor-free survival of heterogeneous tumor normalized to that of a parental-only empty vector control) at varying subpopulation proportions. Each white dot represents a tumor composition for which experiments were performed to determine tumor-free survival in mice. Three out of the four dots at the corner of the ternary plot represents the homogeneous tumor that was used to generate the model. The position of the internal white dot approximates the heterogeneous tumor composition at the end of the experiment in vehicle-treated mice. (B-C) Predicted and actual experimental trajectories of tumor with empty vector control parental/mls/mlt (B) or heterogeneous tumor with parental/shChk2/shBok (C) upon treatment with Vin/SAHA or IRT/CBL in vivo. For heterogeneous tumor (C), the ternary plot was overlay with heat map in (A) showing the therapeutic efficacy comparisons between Vin/SAHA and IRT/CBL. Dotted circle denotes initial tumor composition; solid circle denotes the mean final tumor composition (of pooled lymph nodes per mouse) at relapse following treatment. Predicted trajectories are shown with gray circles and dotted arrows; experimental results of trajectories are shown in white circles and solid arrows.