High-Order Drug Combinations Are Required to Effectively Kill Colorectal Cancer Cells

The advent of targeted therapies for cancer has provided effective treatments for a variety of malignancies, and the study of drug resistance mechanisms has revealed the complexity of tumor biology. A critical challenge is to improve clinical responses to targeted agents, which are often incomplete and/or transient, due either to suboptimal dosing from toxicity concerns (1, 2), the ability of cancers to resist targeted inhibitors via feedback mechanisms, additional oncogenic drivers acting in relevant pathways, or by acquired resistance due to selective pressures (3–7). Drug combinations can potentially overcome these hurdles by countering cellular defense mechanisms, by inhibiting redundant or compensatory pathways, and by targeting emerging resistance (8). In the clinic, combinations of targeted drugs have significantly improved patient survival (9, 10). However, it is not yet clear how many drugs are required and which processes need to be targeted together to cure cancers.

Colorectal cancers, with poor prognoses and complex genetic profiles, provide an ideal test bed for exploring robust cancers and are a major focus of targeted drug discovery. In addition to chemotherapy, patients with metastatic colorectal cancer currently are prescribed 3 FDA-approved targeted therapies: bevacizumab targeting the VEGF (11) and either cetuximab or panitumumab targeting the EGF receptor (EGFR; refs. 12, 13). However, in the case of EGFR-targeted therapy, several downstream pathway aberrations, such as mutations in KRAS or BRAF, have been shown to confer upfront resistance to the treatment and are now used for patient selection. Furthermore, patients who initially respond to EGFR inhibitors often develop resistance within several months. Activating lesions in downstream pathway components (e.g., KRAS) or bypass pathways (e.g., MET) have been found in a large proportion of relapsed tumors (14). Collectively, this suggests that more effective treatments are required to target subsets of KRAS- and BRAF-mutant colorectal cancers, which together make up 50% to 60% of reported cases (15).

In this study, we systematically explored treatments involving up to 6 clinically relevant drugs, targeting oncogenic pathways and key cellular processes in a set of colorectal cancer cell lines representing KRAS and BRAF mutant tumors. We find that under combination treatments, the inhibition of cell growth and activation of apoptosis depend on the genetic context in targeted processes/pathways; the number of drugs (order) used in combination; and in high-order combinations, on the cell's responses to lower-order treatments. Cell models that resist various drug combinations show increased expression of antiapoptotic proteins such as BCL-XL upon treatment. In such cases, inhibition of antiapoptotic proteins was required to achieve apoptotic killing and maximum efficacies in vitro and in vivo. Together, our study presents the first systematic approach to identify high-order drug combinations that can overcome robust models of cancer and suggests specific combinations for the treatment of colorectal cancer to pursue further in preclinical and clinical studies.

All colorectal cancer lines used for this study were obtained from the Novartis/BROAD cancer cell line encyclopedia (Supplementary Table S1; ref. 16). Cell lines were thawed from frozen stocks, cultured in an incubator at 37°C and 5% CO₂, and expanded through ≥1 passage before plating.

Cells were plated in black 384-well microplates with clear bottom (Thermo) in 50 μL media per well using a MultiDrop Combi dispenser (Thermo) at densities between 500 and 1,250 cells per well (Supplementary Table S1) and incubated at 37°C, 5% CO₂ for 24 hours. One plate was counted by microscopy prior treatment (=“Day 0”), and the other plates were treated with 25 nL of 2,000× compound from drug master plates (7-dose points with 1:3 dilution steps, see Supplementary Table S2) using an ATS acoustic dispenser (ECD Biosystems). For pairwise and triple combinations, compound doses were combined at a fixed ratio of 1:1 and 1:1:1, respectively (17). DMSO concentrations were normalized per well. In addition, vehicle (DMSO) and positive controls (staurosporine) were transferred. All treatments were performed in triplicates. One of the replicates received 50 nL of 2 mmol/L CellEvent Caspase-3/7 Green Detection Reagent (Thermo) using a HP D300 dispenser (Tecan). Cells were treated for 72 to 96 hours. Caspase-3/7 activity was imaged every 24 hours. At the end of the treatment, cells were fixed, and their DNA was stained and imaged (18, 19). High-order experiments combining more than 3 drugs were also setup with fixed ratios but only testing 4 dose points. To validate synergies of MDM2 combinations, matrices were tested to assess effects at different ratios of the single agents.

Effects of treatments on apoptosis were calculated from the number of apoptotic nuclei relative to the total number of nuclei, and effects of treatments on cell proliferation were calculated from cell counts relative to DMSO treatment. For each treatment, dose–response curves were fitted and IC₅₀ values calculated using a 4-parameter log-logistic function. Synergy scores were calculated on the basis of normalized cell counts using the highest single agent (HSA) model as a reference (20) and transformed into z-scores to assess the distance of combination effects from a group of noninteracting combinations. Unsupervised, hierarchical clustering of synergy z-scores on the basis of Euclidean distances was used to identify patterns of synergy across cell lines. The most efficacious conditions per treatment were identified calculating maxima of growth inhibition (21) and caspase-3/7 activation across all doses. To validate synergies of MDM2 combinations, the Loewe dose additivity model (22, 23) and isobologram analysis were used to calculate combination indices (CI), which indicate synergy (CI < 1), additivity (CI = 1), or antagonism (CI > 1).

For colony formation assays, cells were plated at densities between 1,000 and 6,000 cells per well in 12-well plates (Costar). Cells were grown for 72 hours before addition of compounds, and treatments/media were refreshed every 48 hours for up to 14 days using an HP D300 dispenser. For extended assays, treatment were either continued as before or washed out (fresh medium every 48 hours). Plates were imaged and analyzed using standard protocols (Supplementary Methods).

For cell-cycle analysis, cells were plated at densities between 1.5 and 3.5 million cells per dish in 10-cm dishes (Corning). Drug treatments were carried out manually after 24-hour incubation. Samples were harvested 24 hours after treatment by collecting both the supernatant and adherent cells, and the cell cycle was analyzed following standard procedures (Supplementary Methods).

For Western blot analyses, cells were plated at densities between 3.5 and 8 million cells per dish in 15-cm dishes (Corning) and grown for 24 hours before manual addition of compounds. Immunoblots were conducted using standard protocols (Supplementary Methods) and using the following primary antibodies at the recommended concentrations: AKT (#9272), pAKT (#4058), BCL-2 (#2870), BCL-xL (#2764), BIM (#2819), EGFR (#4267), pEGFR (#3777), ERK (#4695), pERK (#4370), MCL-1 (#5453), p21 (#2947), cPARP (#9541), PUMA (#4976), all from Cell Signaling Technology; cyclin D1 (sc-718), p27 (sc-528), p53 (sc-126), all from Santa Cruz Biotechnology; β-actin (Ambion, AM4302); MDM2 (CalBiochem, #OP46). The following secondary antibodies were used 1:10,000: HRP goat anti-rabbit (170-5046), HRP goat anti-mouse (170-5047), both from Bio-Rad; IRDye 800CW Goat anti-mouse (Licor, 925-32210).

For expression analysis, cells were plated at densities between 0.25 and 0.75 million cells per well in 6-well plates (Corning). Drug treatments were carried out manually after 24-hour incubation. Samples were collected 10 hours after treatment and processed for qRT-PCR analysis following standard protocols by using probe/primer sets listed in Supplementary Methods.

Tumors were induced by subcutaneous injection of 3 million HCT-116 cells (in 100 μL PBS) in the right flank of female Crl:NU(NCr)-Foxn1nu mice (Charles River Laboratories International). Mice were randomized, and treatments were started when tumors reached 50 to 250 mm³ in volume, for 8 mice per cohort. Tumor volumes (mm³) were monitored 3 times weekly and calculated using the formula: = length × width² × 0.5. CGM097 was dissolved in 0.5% hydroxypropyl methylcellulose, trametinib was dissolved in 1% carboxymethylcellulose containing 0.5% Tween-80% in distilled water (pH7.6-8.0), and navitoclax was dissolved in Microemulsion pre-concentrate 5. Drugs were dosed orally using 5 to 10 ml/kg. CGM097 was administered 3 times a week (3 qw) at 100 mg/kg; trametinib and navitoclax were administered daily (q24h) at 0.3 and 100 mg/kg, respectively. The combination dosing schedule and dosage were the same as the single reagents. For statistical analysis, tumor volumes at each time point were normalized to the volume before the start of the treatment to obtain “percent change of tumor volume”.

To identify effective drug combinations, we used high-throughput imaging to measure effects on cell viability and apoptosis for single drugs as well as pairwise and triple combinations in colorectal cancer cell lines (Fig. 1A). We focused our study on cell lines with genetic alterations in the RAS/MAPK pathway (KRAS, BRAF), subsets of which also harbored lesions in the PI3K/AKT, and/or p53 pathways, resulting in 23 cell lines total (Supplementary Table S1; ref. 16). We selected 20 clinically relevant drugs targeting processes critical for colorectal cancer tumorigenesis and progression for screening (Supplementary Table S2). Specifically, these included key signaling pathways (RAS/MAPK and PI3K/AKT, hereafter referred to as “growth pathways”), signaling through receptor tyrosine kinases (RTK), apoptosis and cell cycle, and standard-of-care (SOC) cytotoxic agents.

Two distinct screens were performed. Since the majority of cell lines had alterations in the RAS/MAPK pathway, we first tested effects of pairwise drug combinations interrogating growth and RTK signaling. In the second screen, we combined the most effective single agents and drug pairs from the first screen with drugs targeting downstream biological processes and systematically assessed pairwise and triple combinations. Each treatment was tested in seven doses, and for combination treatments, the underlying single agents were combined at a fixed ratio (17). We used fluorescence microscopy to count caspase-3/7–positive cells live every 24 hours as a proxy for apoptosis and the total number of cells as proxy for cell viability (Supplementary Fig. S1A).

We used cell counts to calculate synergy scores on the basis of the HSA model (20). This identified pairwise combinations that exceeded the effect of the highest single agent or in the case of triple combinations those that exceeded the highest pairwise combination effects (Supplementary Fig. S1B). From these synergy scores, we calculated z-scores relative to combination effects from a group of noninteracting combinations and focused further analyses on combinations with z-scores ≥ 3 (Supplementary Figs. S1C and S1D and S2A–S2C).

To assess the extent to which treatments could deplete cells, we compared cell counts before the start (“day 0”) and at the end of each treatment. This is represented by the growth inhibition (GI) scale (21), which ranges from 0 to 200 and quantitatively differentiates treatments with no effect (GI = 0), cytostatic treatments (GI = 100, equals “day 0”), and cytotoxic treatments that deplete the number of cells below the “day 0” count (GI > 100; Supplementary Figs. S1A and S2D).

In general, single agents showed limited efficacy and only about 10% were cytotoxic and activated caspase-3/7 in more than 25% of the cells (Fig. 1B and Supplementary Fig. S3A–S3C). This was seen despite evidence of strong target inhibition (Supplementary Fig. S3D and S3E) and selectivity of compounds for particular genetic subsets (Supplementary Fig. S3F and Supplementary Table S3). Interestingly, single-agent treatments showed stronger GI with increasing caspase-3/7 activation (Fig. 1B). For example, inhibition of MTOR in LIM2405 did not activate caspase-3/7 and the maximum GI observed was 80, whereas inhibition of ERK activated caspase-3/7 in more than 50% of the cells and led to a GI of 145 (Fig. 1C). We found a positive correlation of GI and caspase-3/7 activation across all treatments, including combinations, in our study (Fig. 1D and Supplementary Fig. S4). However, this relationship was not absolute (Pearson correlation ∼ 0.7), indicating that viability and apoptosis assays provide some complementary information about treatment effects. This suggests that other methods of cell death may occur in these cells and that activation of apoptosis can be decoupled from effects on cell proliferation.

To address the limited efficacy of single agents, we systematically explored pairwise drug combinations targeting key nodes of growth and RTK signaling (13 compounds, Supplementary Table S2) in 23 cell lines (Supplementary Table S4).

Overall, the combinations were more effective than the single-agent treatments and 54% (967 of 1,794) of the drug pairs were synergistic with z-scores ≥ 3. We used hierarchical clustering to identify unique patterns of interactions across the cell lines (Fig. 2A and Supplementary Fig. S5A). Drug combinations that co-targeted the RAS/MAPK and PI3K/AKT pathways (horizontal combinations) and combinations that targeted each of these pathways twice (vertical combinations) showed strong synergies across the majority of cell lines (clusters 1 and 5 in Fig. 2A). Cytotoxic effects and caspase-3/7 activation were seen primarily when the RAS/MAPK pathway was targeted as part of the combination (Fig. 2B and Supplementary Fig. S5B–S5D). Interestingly, we found multiple horizontal combinations to be similarly efficacious despite targeting distinct nodes within each pathway (Fig. 2C and D and Supplementary Fig. S6).

RTK inhibitor combinations (clusters 2–4) showed more cell line–specific interactions, except for linsitinib (IGF1R/IR), which grouped with combinations targeting RAS/MAPK and PI3K/AKT pathways and provided combination benefits in almost all of the lines tested (193 of 276 combinations with z-score ≥ 3; Fig. 2E and Supplementary Fig. S7A). Instead, lapatinib (HER2) and erlotinib (EGFR) combinations were most synergistic in BRAF/KRAS wild-type and in some KRAS-mutant models but overall provided synergy less often (142 of 276 and 113 of 276 combinations, respectively). Moreover, their synergy scores and maximum GIs were positively correlated across cell lines (Pearson correlation of 0.85 and 0.91, respectively), which was consistent with both drugs likely to inhibiting EGFR and HER2 (24, 25). Combinations with PF-04217903 (MET) were almost exclusive to RKO and LoVo cells and were very rare (19 of 276). Interestingly, RKO overexpresses the MET ligand HGF (16), and LoVo expresses a noncleaved, constitutively active form of MET (26), which likely explains their sensitivity to MET inhibitor combinations and highlights the specificity of our screening approach. None of the RTK inhibitors preferentially interacted with inhibitors of RAS/MAPK or PI3K/AKT signaling (Fig. 2F and Supplementary Fig. S7B), and strong cytotoxic effects required targeting of the RAS/MAPK pathway in combination (Fig. 2G).

Unexpected, selective synergies of the combinations tested were rare. Vertical combinations targeting PI3K/AKT signaling (e.g., combined PIK3CA and AKT inhibition) had significantly stronger synergies in BRAF mutants when compared to KRAS mutants, and horizontal combinations targeting PIK3CA were more synergistic in PIK3CA mutants compared to wild type cell lines (Supplementary Table S3).

Despite the synergies seen in many of the combinations tested, only few led to cytotoxic effects (23% or 222 of 967 with maximum GI > 100) and activation of apoptosis (17% or 169 of 967 with maximum caspase-3/7 activation ≥ 25%). This observation suggests that targeting additional processes, or combinations of more than 2 molecules, are required for overall better efficacy.

We systematically tested all pairwise and triple combinations of selected RAS/MAPK pathway inhibitors, PI3K/AKT pathway inhibitors, and RTK inhibitors with compounds targeting downstream, cancer-relevant processes (Fig. 1A). We focused on targets with impact on cellular stress, hypothesizing their perturbation would result in increased cell death by altering key mediators of apoptotic responses (27) and selected 2 agents targeting the apoptosis pathway (BCL-2 family inhibitor navitoclax, hereafter referred to as BCL-2/XL inhibitor, and IAP inhibitor LCL161), 2 agents targeting cell cycle (MDM2 inhibitor CGM097 and CDK4/6 inhibitor ribociclib), and 3 SOC agents (5-FU, SN-38, paclitaxel; Supplementary Tables S2 and S4).

Of 1,170 drug pairs tested, 489 (∼42%) had synergy z-scores ≥ 3 and showed target-dependent interaction patterns across cell lines (Fig. 3A and Supplementary Fig. S8A). Distributions of synergy scores indicated that of the additional agents tested, those that targeted cell cycle had a higher fraction of synergistic combinations than those targeting apoptosis and SOC agents (Fig. 3B). Combinations of the CDK4/6 inhibitor ribociclib with growth pathway inhibitors were broadly synergistic across cell lines (Fig. 3A). Surprisingly, its combination with erlotinib (EGFR cluster in Fig. 3A) showed selective synergy in KRAS mutants compared with BRAF mutants. Furthermore, ribociclib selectively synergized with the MDM2 inhibitor CGM097 in p53 wild-type lines (Fig. 3A and Supplementary Table S3). CGM097 itself had several strong synergies with growth pathway inhibitors, including trametinib, AZD-6044, and navitoclax. While combinations of ribociclib rarely induced apoptosis or GI levels in excess of 100, combinations of CGM097 often depleted the number of cells below input and moderately activated apoptosis (Fig. 3C, Supplementary Fig. S8B–S8E). Combinations involving navitoclax or the SOC paclitaxel induced apoptosis strongly and showed cytotoxic GI (Supplementary Fig. S8F–S8I). However, paclitaxel had strong single-agent effects already, whereas navitoclax was not effective as single agent in most of the cell lines (Supplementary Fig. S3A and S3B) and subsequently showed stronger synergies.

Generally, effects of combinations were variable between cell lines and did not obviously correspond to a cell line's driver mutations, suggesting the existence of additional factors that determine synergy and efficacy. For example, combinations of alpelisib (PIK3CA) with CGM097, navitoclax, or paclitaxel were synergistic and cytotoxic in GP2d cells but lacked combination benefits in HCT-116 cells, despite their similar genetic alterations (Fig. 3D and Supplementary Fig. S8E, S8G, and S8I). These variations were seen for many drug pairs with no obvious relation to the cells' genetics (Fig. 3E).

Triple combinations proved more effective at killing than the best pairwise synergies (Fig. 4A). The analysis of 4,290 triple combinations identified 600 synergies with z-scores ≥ 3 (∼14%), and of 352 distinct triple combinations in this screen, 221 were scored as “hits” in at least one cell line (Supplementary Table S4). A systematic comparison of synergies between drug pairs and drug triples showed a trend of higher synergy scores for triple combinations with increasing synergy scores of the underlying pairwise combinations (Fig. 4B and C and Supplementary Fig. S9), which was consistent with previous observations (28).

Sufficient growth pathway inhibition was crucial for potent triple combinations (Fig. 4D–F). Most cell lines showed synergistically enhanced efficacies when growth pathways were inhibited by triple combinations, including the combination targeting MEK (trametinib), RAF (RAF265 or dabrafenib), and PIK3CA (alpelisib; Fig. 4D and E and Supplementary Figs. S10 and S11). Combinations targeting growth pathways and CDK4/6 were the most broadly synergistic triple combinations identified. The lack of apoptosis activation suggested that cell-cycle arrest induced by CDK4/6 inhibition was the dominant antiproliferative mechanism in these combinations (Supplementary Fig. S12A). In contrast, triple combinations involving navitoclax led to strong caspase-3/7 activation and cytotoxic GI levels (Supplementary Fig. S12B). Some of the KRAS-mutant models that did not show strong responses to the drug pair of navitoclax with trametinib activated caspase-3/7 more strongly when another inhibitor suppressing growth signaling was used, such as alpelisib in HCT-116 cells (Fig. 4F). The MDM2 + MEK inhibitor combination, which yielded one of the most consistent pairwise synergies (average z-score of 11 ± 3 in 5 p53 wild-type models), showed enhanced GI and apoptosis when either PIK3CA or BCL-2/XL were also inhibited (Fig. 4E and F and Supplementary Fig. S12D–S12N). This observation prompted us to further investigate this combination.

Roughly 50% of patients with KRAS- and BRAF-mutant colorectal cancer are wild-type for the p53 gene (∼30% of total patients with colorectal cancer; ref. 29) and may theoretically benefit from treatments with MDM2 inhibitors or combinations thereof. The pairwise drug combination screen indicated that combined inhibition of MDM2 (CGM097) and MEK (trametinib) was superior over each of the single agents, but the activation of apoptosis was moderate (average of 18% ± 13%), and the GI was not indicative of absolute cell killing (average of 133 ± 23).

Exploring similar combinations across 8 cell lines validated the synergy and efficacy of this drug pair. We tested CGM097 in combination with MEK, ERK, and BRAF inhibitors, and isobologram analysis showed synergy over Loewe dose additivity (30, 31) under different drug ratios (Fig. 5A and Supplementary Fig. S13A). For example, CIs for CGM097 + trametinib (MDM2 + MEK) indicated that only half of the total compound dose was required in combination to achieve the same effect as each of the single agents (average CI of 0.43 ± 0.18). Furthermore, MDM2 + MEK inhibition prevented outgrowth of clones significantly better than each of the single agents in colony formation assays (Fig. 5B and C and Supplementary Fig. S13B).

The combination prevented cell-cycle progression. We performed fluorescence-activated cell-sorting (FACS) analysis after MDM2 + MEK inhibitor treatment (Fig. 5D). MDM2 inhibition depleted cells from S-phase. MEK inhibition in the majority of models resulted in increased G₁ populations. In the combination, S-phase depletion was retained and 5 of 8 models showed increased sub-G₁ populations, suggesting cell death (Supplementary Fig. S13C).

The combination modulated regulators of cell cycle and apoptosis. In quantitative real-time (qRT) PCR analyses of cell-cycle regulators, MDM2 inhibition increased expression of p21 (Fig. 5E and Supplementary Fig. S13D), whereas MEK inhibition decreased the expression of cyclin D1 in all models and increased expression of p27 in only some of the models. Collectively, these results likely explain the cell-cycle arrest seen in FACS. Analyzing apoptosis regulators, MDM2 inhibition increased mRNA expression of proapoptotic PUMA and BAX, whereas MEK inhibition increased expression of proapoptotic NOXA-1, PUMA, and BMF. Generally, combined inhibition of MDM2 and MEK did not yield strong combinatorial effects on the cell-cycle and apoptosis regulators tested, with the possible exception of PUMA and NOXA-1. Western blotting (Fig. 5F and Supplementary Fig. S13E) confirmed the effects of treatment on cell cycle and apoptosis. In addition, MEK inhibition led to higher levels of proapoptotic BIM. Consistent with the caspase-3/7 activation seen in the screen and the increased sub-G₁ population seen in FACS analysis, for 6/8 models, we saw increased levels of cleaved PARP after combination treatment compared with single-agent treatments, which indicates activation of apoptosis. Overall, our data suggest a model where both single agents regulate complementary sets of cell-cycle proteins to induce arrest and proapoptotic proteins to induce cell death (Fig. 5G and Supplementary Table S5). The precise mechanism of action, for example which specific factors are crucial for these mechanisms, remains to be determined.

Despite consistent synergy, however, apoptosis activation produced mostly incomplete cell killing, both in the experiment's GI levels and in the colony formation assays. When the most synergistic treatments (arrowheads in Fig. 5B and C) were extended or washed out for 10 days, we found that some lines (e.g., HCT-116 and LS-180) started growing back under sustained treatment, and all lines grew back when the treatment was discontinued (Fig. 6A and B), confirming incomplete cell killing by the initial treatment.

The triple combination screen identified navitoclax to strongly induce apoptosis in the background of MDM2+MEK inhibition, which pointed toward a role of antiapoptotic proteins for incomplete cell killing. Monitoring MCL-1, BCL-XL, and BCL-2 after MDM2 + MEK inhibition in vitro by Western blotting showed increased expression of BCL-XL in all models compared with controls (Fig. 6C). Assuming that BCL-XL, at least in part, protected surviving cells, this suggested that navitoclax could improve responses even when added sequentially. In vitro, 4 of 6 models showed improved GI and caspase-3/7 activation when navitoclax was added sequentially to the MDM2 + MEK inhibitor combination (Fig. 6D and Supplementary Fig. S14).

The triple combination's increased effectiveness was confirmed in vivo. In a xenograft model of HCT-116 cells, the combination treatment using clinically relevant doses of CGM097 and trametinib led to stable disease, whereas the concomitant triple combination with navitoclax led to marked tumor regressions and significantly smaller tumor sizes compared with the drug pair (P < 0.01, Fig. 6E and F). Strikingly, sequential addition of navitoclax to the MDM2 + MEK pair after 9 days led to marked tumor regressions that were indistinguishable from the concomitant treatment (Fig. 6G) and significantly more effective than the best pair (trametinib + navitoclax, P < 0.05). However, this 3-way combination still yielded incomplete killing and in vitro showed increased expression of MCL-1 (Fig. 6H).

To investigate the importance of antiapoptotic signaling for complete killing of robust colorectal cancer cell lines, we explored higher order combinations. Across the synergistic pairwise and triple combinations in the screen, we found that a subset of “most resistant” cell lines showed only limited maximum responses to the majority of treatments (Fig. 7A).

We chose some of the most synergistic combinations in the 2 most “resistant” BRAF- and KRAS-mutant lines (RKO, HT-29, and DLD-1, LS-180, respectively) and monitored antiapoptotic proteins under treatment. Remarkably, in DLD-1 cells, all pairwise combinations targeting MEK, PIK3CA, and EGFR as well as the resulting triple combination showed higher expression of BCL-XL and BCL-2 than controls (Fig. 7B and Supplementary Fig. S15A). This was also seen for combinations targeting BRAF, MDM2, MET, and PIK3CA in the BRAF-mutant line RKO (Fig. 7C, Supplementary Fig. S15B), as well as in additional lines (Supplementary Fig. S15C and S15D). This suggested that additional inhibition of antiapoptotic proteins was required to achieve adequate killing.

We systematically investigated combinations of up to 6 drugs and combined the selected synergistic treatments with inhibitors that could potentially enhance their efficacies including ribociclib (CDK4/6), CGM097 (MDM2), and navitoclax (BCL-2/XL). For the 6 drugs, we covered all possible submixtures up to the full sixth-order combination and scored maximum GI and apoptosis and differential synergy relative to the best included submixture (Supplementary Table S6; ref. 17).

Synergy and maximum killing effects tended to saturate at fourth order, with no further benefit as more drugs were added (Fig. 7D and E). For example, in DLD-1 cells, we found the most efficacious and synergistic combination of 4 drugs targeted MEK, PIK3CA, EGFR, and BCL-2/XL (Fig. 7F). Inhibition of BCL-2/XL was required for activation of apoptosis and cytotoxic growth inhibition in all submixtures. In RKO cells, inhibition of BCL-2/XL on top of BRAF, MET, and MDM2 inhibition provided the most efficacious and synergistic combination of 4 drugs in this experiment (Fig. 7G and Supplementary Fig. S15B). Similar benefits of BCL-2/XL inhibition at high order were seen in additional models (Supplementary Fig. S15E and S15F).

The increasing arsenal of agents targeting oncogenic signaling opens many new avenues for drug combinations to overcome adaptive and acquired drug resistance. Recent studies demonstrated the power of in vitro screening to identify combinations that can overcome limited responses to single-agent targeted treatments (22, 23). Here, we used a combinatorial, high-order drug screening approach—to our knowledge the first systematic study of this kind in cancer cells—that assessed the potential of treatments to inhibit cell growth and to induce apoptosis. The screen identified novel drug combinations with potential for high efficacy in colorectal cancer including agents targeting RAS/MAPK and PI3K/AKT pathways combined with agents targeting CDK4/6, MDM2, or BCL-2 family proteins.

Our results indicate that to achieve optimal efficacy, it is critical to target essential growth mechanisms, like RAS/MAPK signaling. This is concordant with recent clinical findings in BRAF-mutant colorectal cancer that highlight the dependence of these cancers on sustained MAPK signaling (32). Pairwise combinations targeting RAS/MAPK and PI3K/AKT pathways resulted in synergies broadly across cell lines. Interestingly, the resulting maximum effects were largely independent of the targeted node in either of the pathways, which could be of interest in clinical settings where targeting alternative nodes could provide better tolerability. This finding might be reflective of extensive cross-pathway feedback control between both signaling cascades (33, 34) and could also result from pathway reactivation mediated by signaling through upstream RTKs (35–37).

Surprisingly, most combinations targeting growth signaling could not adequately kill cells. One of the drug pairs tested, targeting MEK and PI3K, has been shown preclinically with antitumor activity in various cancer models and genotypes, but clinical trials so far have only shown modest activity of the combination (38). Together, this suggests that other cellular processes need to be targeted in combination and that dual inhibition therapy, even when targeting the tumor's major genetic alterations, might not be sufficient for adequate killing.

Testing a broader range of drug pairs identified interesting synergy patterns for ribociclib (CDK4/6), CGM097 (MDM2), and navitoclax (BCL-2/XL). The CDK4/6 inhibitor provided strong synergies when combined with inhibitors targeting growth signaling and MDM2 but was not able to induce apoptosis. This is in agreement with recent reports of CDK4/6 combinations with MEK inhibitors, which exert their effects mainly through cell-cycle arrest and lead to tumor stasis (39, 40). Inhibitors targeting the MDM2/p53 interaction stabilize p53 and lead to p53-induced cell-cycle arrest and/or apoptosis in p53 wild-type cancers and are now in clinical trials (41–43). MDM2 inhibitors have been shown to cooperate with inhibition of oncogenic signaling pathways (44, 45). We found several potent and synergistic combinations in our screen, including the combined targeting of MDM2 and MEK, which provides a promising candidate for further investigation. In contrast to combinations with CDK4/6 and MDM2 inhibitors, drug pairs involving the BCL-2 family inhibitor navitoclax were synergistic in fewer cell lines but strongly activated apoptosis, which was expected from previous findings (46, 47). However, most pairwise combinations tested showed variable effects on growth and apoptosis across cell lines, even in models with the same oncogenic drivers. This observation has also been made in a recent combination screen in melanoma cells (48) and suggests different capabilities of cell models to buffer perturbations by redundancies or other mechanisms, which could potentially be overcome with additional molecules.

Triple combinations could enhance both growth inhibition and apoptosis when compared with their underlying drug pairs and in some cell lines were required to tip the balance from cytostatic to cytotoxic effects. Dual growth pathway inhibition provided a good “base” for most of the third agents tested, including CDK4/6 and MDM2 inhibitors. This might result from the critical role of growth pathways in regulating pro- and antiapoptotic factors being crucial to sufficiently prime cells for apoptosis (27, 49).

Our validation studies suggest that the balance of pro- and antiapoptotic regulators underlies many observed synergies. The combination targeting MDM2 and MEK could induce different proapoptotic proteins through activating p53 signaling and inhibiting RAS/MAPK signaling, respectively. However, the combination was only moderately cytotoxic and cells that survived sustained treatments show increased expression of antiapoptotic BCL-XL. Sequential addition of navitoclax enhanced in-vitro cell killing and led to marked tumor regressions in vivo, suggesting that BCL-XL, at least in part, protected cells from apoptosis. However, tumor regressions remained incomplete even with triple combinations, perhaps due to increased expression of other antiapoptotic proteins such as MCL-1. It would be interesting to study whether cell populations with increased expression of antiapoptotic proteins pre-exist, whether their expression is adaptively induced upon treatment, or whether both mechanisms are important to protect cells from apoptotic cell death.

Robust colorectal cancer cell lines that resisted most pairwise and triple combinations tested also showed increased expression of antiapoptotic proteins. In these cases, fourth-order combinations were required to reach apoptotic killing, which targeted for example RAS/MAPK and PI3K/AKT signaling, an upstream RTK, and at least one downstream target (MDM2 or BCL-2/XL). To control proliferation and induce apoptosis in such models, it seems essential to target cellular signaling at several levels. This is similar to findings from studying the dynamics of cancer in response to targeted combination therapy (50), and suggests that these cell lines exhibit greater functional complexity, with access to buffering via multiple growth signaling pathways (17). The order and exact choice of inhibitors depends on the driver genetics and treatment responses of the cells.

Our primary screening results provide evidence for the efficacy of high-order combinations to overcome robust models of colorectal cancer; however, they do not address the important concern of synergistic toxicity. Our results should be considered preliminary and tested for toxicities in “normal” tissues such as in vivo systems or tissue explants to further prioritize on the basis of tolerability. However, we do show in vivo that the dimension of combinations can be reduced through sequential dosing regimens, which may decrease toxicity while retaining efficacy. This concept requires a good understanding of the mechanisms of synergy and might not be applicable for all high-order combinations. With the number of targeted agents being developed and the growing understanding of alterations and resistance mechanisms of cancers, we expect strategies to achieve high-order combinations for the treatment of cancer to be an active area of future research.

R. Schlegel has ownership interest (including patents) in Novartis and is a consultant/advisory board member of Third Rock Ventures. E. Morris is employed as a Research Scientist at Novartis. J. Greshock is employed as a Sr. Director, Bioinformatics at Neon Therapeutics and is a consultant/advisory board member of Third Rock Ventures. L.A. Garraway is a consultant at Foundation Medicine, Novartis, Boehringer Ingelheim, Third Rock; reports receiving a commercial research grant from Novartis and other commercial research support from BMS, Merck, Astellas; has ownership interest (including patents) in Foundation Medicine; and is a consultant/advisory board member of Warp Drive. No potential conflicts of interest were disclosed by the other authors.

Conception and design: T. Horn, S. Ferretti, S. Ho, D. Porter, E. Morris, S. Jeay, L.A. Garraway, G. Caponigro, J. Lehár

Development of methodology: T. Horn, A. Tam, A. Farsidjani, E. Halilovic, J. Lehár

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): T. Horn, S. Ferretti, N. Ebel, F. Harbinski, A. Farsidjani, M. Zubrowski

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): T. Horn, S. Ferretti, W.R. Sellers, J. Wuerthner, S. Jeay, E. Halilovic, L.A. Garraway, J. Lehár

Writing, review, and/or revision of the manuscript: T. Horn, W.R. Sellers, D. Porter, E. Morris, J. Wuerthner, S. Jeay, J. Greshock, E. Halilovic, G. Caponigro, J. Lehár

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): T. Horn, A. Tam, F. Harbinski, J. Wuerthner

Study supervision: S. Ferretti, R. Schlegel, J. Greshock, G. Caponigro, J. Lehár

We thank S. Silver, D. Stuart, K. Venkatesan, F. Stegmeier, A. Huang, and R. McDonald for critical discussion of the article and M. Stump for technical assistance for screening. We also thank A. Szilvasi and A. Ho for assistance with flow cytometry.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.