Exploiting Allosteric Properties of RAF and MEK Inhibitors to Target Therapy-Resistant Tumors Driven by Oncogenic BRAF Signaling

RAF kinases BRAF, CRAF, and ARAF signal through their substrate MEK to activate MAPK signaling and promote cell proliferation and survival. Mutationally activated BRAF^V600E kinase drives growth of about 8% of human tumors (1, 2), most commonly melanomas and colorectal cancers, and three small-molecule RAF inhibitors [vemurafenib (VEM), dabrafenib (DAB), and encorafenib (ENC)] are now FDA-approved drugs (3). Unlike most kinase inhibitors that bind and inhibit their target in all cells, these RAF inhibitors selectively inhibit the mutationally activated form of the protein (BRAF^V600E) in tumors but not wild-type BRAF (BRAF^WT) in normal tissue. A model to explain this observation has been put forward by us and others, in which BRAF^WT signals as an obligatory dimer, but BRAF^V600E is able to signal as a monomer, and these RAF inhibitors preferentially bind and inhibit monomeric over dimeric RAF (4–6). Subsequent studies revealed the structural basis of this property as negative allostery for inhibitor binding to the second protomer in the BRAF dimer (7, 8). Nonetheless, although the inability of this class of RAF inhibitors to inhibit dimeric BRAF is the basis of their increased therapeutic index, it is also responsible for development of both acquired and adaptive drug resistance. In adaptive resistance, relief of negative feedback upon MAPK pathway inhibition results in rapid formation of RAF dimers and consequent resistance to RAF inhibitors (9, 10). To address these limitations of the monomer-selective RAF inhibitors, RAF inhibitors that are equipotent inhibitors of RAF dimers and monomers have entered preclinical and clinical development (11, 12). These drugs are predicted to overcome resistance due to RAF dimerization. However, as they are also potent inhibitors of dimeric BRAF^WT, they are also predicted to cause on-target toxicities in normal tissue at the high doses required for potent antitumor effect. This may account for the fact that none of these compounds have been successful in the clinic yet. Thus, despite significant progress, there is a pressing need to develop more effective therapeutic strategies for most patients with BRAF^V600E tumors.

A large number of MEK inhibitors have also entered preclinical and clinical development, and four of them [trametinib (TRAM), selumetinib, cobimetinib (COB), and binimetinib (BIN)] are FDA-approved drugs. The impact of allosteric properties on the antitumor effectiveness of MEK inhibitors in different cellular contexts has also drawn a great deal of interest (13–15). However, which MEK inhibitors should be optimally used in combinatorial regimens with RAF inhibitors in BRAF^V600E tumors remains elusive.

Here, we undertook a systematic approach to biochemically classify and characterize RAF and MEK inhibitors in preclinical or clinical development. We identified and characterized a new class of RAF inhibitors that are selective for dimeric RAF and used them to gain new mechanistic insight into conformational states adopted by BRAF^V600E in its native state. We further used this knowledge to design a more effective combinatorial strategy for BRAF^V600E tumors, with the goal of maximally suppressing MAPK signaling in the tumor while retaining a broad therapeutic index.

Protein kinases exist in an equilibrium of active and inactive states, where the switch between these states involves movements of two conserved structural motifs: the Asp–Phe–Gly (DFG) motif and the αC-helix. Each of these motifs can adopt an “IN” or “OUT” conformation, where both motifs in the “IN” conformation (αC-OUT/DFG-IN) are associated with the catalytically active state. Small-molecule inhibitors of kinase can be grouped by the kinase conformations they recognize (16). For example, the type I kinase inhibitor gefitinib (GEF) binds the ATP binding site in the active αC-IN/DFG-IN state (“CIDI”), while the type I1/2 inhibitor VEM and the type II inhibitor sorafenib (SOR) bind to the inactive αC-OUT/DFG-IN and αC-IN/DFG-OUT states (“CODI” and “CIDO,” respectively; ref. 17).

We previously characterized structurally and biochemically two classes of RAF inhibitors. One class of CODI inhibitors selectively inhibits monomeric compared with dimeric BRAF^V600E (RAF monomer selective) and includes the current clinical inhibitors VEM and DAB. The other class (including both CIDO and CIDI inhibitors) has potent inhibitors of both monomeric and dimeric BRAF^V600E due to similar affinity for inhibitor binding to both RAF protomers (7, 8). To identify RAF inhibitors with potentially novel properties, we used an approach that enabled us to determine the relative potency of each RAF inhibitor in inhibiting monomeric versus dimeric BRAF^V600E in the endogenous setting (7, 8): SKMEL239 Parental (PAR; i.e., endogenously expressing full-length monomeric BRAF^V600E) and C3, a clone of RAF inhibitor–resistant SKMEL239 cells that endogenously expresses a splice variant of constitutively dimerized BRAF^V600E (5).

We first assayed the most recently FDA-approved RAF inhibitor ENC (18), which, as expected, and similar to the other clinical CODI RAF inhibitors, VEM (19) and DAB (20), showed more potent inhibition of MAPK signaling in PAR compared with C3 cells (Fig. 1A). Furthermore, other CIDO RAF inhibitors, such as AZ628, TAK632 (TAK), LY3009120, and BGB283, showed similar potency in the two settings, confirming their similar potency for BRAF^V600E monomers and dimers (7, 8, 12). Surprisingly, we identified four CIDO RAF inhibitors [SOR (21), regorafenib (REG; ref. 22), RAF709 (23), and LXH254 (24)] that suppressed MAPK signaling more potently in C3 cells compared with PAR (Fig. 1A), suggesting that they might be more potent inhibitors of dimeric than monomeric BRAF^V600E. The increased potency of these compounds in suppressing MAPK signaling in cells expressing dimeric BRAF^V600E was also confirmed in another pair of Parental (M397-P) and RAF inhibitor–resistant cells that express a different BRAF^V600E splice variant (M397-R; ref. 25; Supplementary Fig. S1A).

To assess whether these compounds inhibit more potently the dimeric compared with the monomeric form of BRAF^V600E in an isogenic system, we used a chemically induced dimerization (CID) system by ectopically expressing BRAF^V600E engineered to dimerize (BRAF^V600E–Dmr) upon treatment with the ligand AP20187 (AP; Fig. 1B). To avoid the effects of baseline RAS–GTP levels, we ectopically expressed BRAF^V600E–Dmr in PC9 cells, in which RAS signaling is under the control of mutant EGFR, and it can thus be suppressed using the EGFR inhibitor GEF (7). Activated MAPK signaling in GEF-treated PC9 cells was inhibited by either VEM or DAB more potently in the absence than in the presence of AP, consistent with these RAF inhibitors being more potent inhibitors of monomeric than dimeric BRAF^V600E. In contrast, REG and LXH254 suppressed MAPK more potently in the presence than in the absence of AP, confirming that their increased potency toward the splice variants of BRAF^V600E is a consequence of their increased potency toward dimeric compared with monomeric BRAF^V600E (Fig. 1C; Supplementary Fig. S1B).

Cellular thermal shift assay (CETSA) enables monitoring changes in the thermal stability of proteins upon inhibitor engagement in intact cells (26, 27). Treatment of PAR cells with VEM resulted in a shift in the protein's melting temperature (Tm), indicating inhibitor binding to full-length BRAF^V600E (i.e., BRAF^V600E monomer), whereas treatment with REG did not affect the Tm of BRAF^V600E. In contrast, treatment of C3 cells with REG resulted in a shift in the Tm, indicating binding of inhibitor with the BRAF^V600E splice variant (i.e., BRAF^V600E dimer), whereas treatment with VEM did not affect the protein's Tm (Fig. 1D and E). Together, these experiments show that the two RAF inhibitors VEM and REG preferentially bind monomeric or dimeric BRAF^V600E in cells, respectively.

Examination of available crystal structures of the RAF dimer–selective inhibitors SOR (PDB:1UWH) and RAF709 (PDB:5VAM) in complex with the catalytic domain of BRAF did not reveal any major differences in their mode of binding compared with other CIDO RAF inhibitors (7) that are equipotent for monomeric and dimeric BRAF^V600E, such as TAK (PDB:4KSP) or LY3009120 (PDB:5C9C). Thus, to identify the basis of the apparent biochemical difference between the two classes of compounds, we employed molecular dynamic (MD) simulations.

We performed 15-μs simulations of seven BRAF inhibitors, each bound to monomeric and dimeric BRAF, with three independent simulations for each condition (Supplementary Table S1). For each inhibitor, simulations were initiated from the poses observed in cocrystal structures (when available) or from a docked pose modeled based on crystal structures of structurally similar compounds (Supplementary Table S1). We analyzed ligand stability using the root mean square deviation (RMSD) of ligand-heavy atoms with respect to the initial pose as an indication of binding strength to a receptor conformation.

In these simulations, all seven ligands remained bound to dimeric BRAF with low RMSD compared with the initial pose, with the overall RMSD ranging from 0.5 to 2 Å. This was not the case for all ligands in the simulations with monomeric BRAF: For the equipotent inhibitors (TAK, AZ628, and LY3009120), the RMSD range was comparable to the RMSD in the simulations with dimeric BRAF; for the dimer-selective compounds (REG, SOR, RAF709, and LXH254), we observed significantly larger ligand RMSD in at least two out of the three independent simulations (Supplementary Videos S1–S4). To understand the structural basis for the differences in ligand stability among the monomeric BRAF simulations, we studied the dynamics of the protein structural elements that are involved in coordinating ligand binding in the specificity pocket (circled in red in Fig. 2A and B). We found that the αC-helix, which sits directly above the specificity pocket, sampled a variety of orientations in monomeric BRAF both in its apo state and when bound to the dimer-selective BRAF inhibitors. In contrast, the motion of the αC-helix was restricted in the inhibitor-bound BRAF dimers (Supplementary Fig. S2A).

On the basis of these observations, we hypothesized that the stable engagement of the specificity pocket is correlated with dimer selectivity for CIDO (type II) RAF inhibitors. These differential specificity pocket dynamics hold for all seven inhibitors studied retrospectively. To provide further evidence that ligand stability is correlated with the motion of the αC-helix, we also simulated REG, a compound we found experimentally to be RAF dimer selective (Fig. 1A and C), with a β3-αC deletion mutant (PDB:5HI2; ref. 28). This mutation locks the αC-helix in a dimer-like “IN” state, which we hypothesized would maintain the low RMSD observed in the simulations of REG-bound monomeric BRAF^WT. As expected, REG was comparably stable in the monomeric and dimeric mutant constructs (Fig. 2C; Supplementary Video S5).

We further applied our simulation method to three additional compounds [belvarafenib, BGB-283 (29), and ponatinib (30)], for which binding selectivity for the RAF dimer had not been determined experimentally, using the same computational protocol we used in the retrospective studies. In our simulations, belvarafenib behaved similarly to the RAF dimer–selective compounds: It had a large RMSD in simulations with monomeric but not dimeric BRAF. BGB-283 showed comparable RMSD in simulations with the BRAF monomer and dimer, indicating equal potency toward each. The ponatinib simulations, however, were less conclusive: On one hand, ponatinib had comparable RMSDs in monomer and dimer simulations, consistent with the behavior of the equipotent compounds. On the other hand, the BRAF αC-helix remained highly dynamic in the monomeric simulation and was not stably engaged, behavior that was observed in simulations of the RAF dimer–selective compounds (Supplementary Fig. S2A and S2B). The dimer selectivity or equipotency of ponatinib is thus unclear based on the MD simulations alone.

We then experimentally assessed the potencies of belvarafenib, BGB-283, and ponatinib in inhibiting monomeric versus dimeric BRAF^V600E using both the PAR/C3 cell approach and the CID system. In accordance with our predictions based on the MD simulations, belvarafenib was found to be a more potent inhibitor of dimeric RAF, whereas BGB-283 was not (Supplementary Fig. S2C and S2D). Ponatinib also showed selectivity toward dimeric RAF, consistent with the BRAF αC-helix remaining highly dynamic in the monomeric simulation when bound to the compound. We have also showed recently that ponatinib stabilizes the αC-helix of BRAF in a unique position compared with other RAF inhibitors (31), which may account for its distinct features. The discrepancy between this finding and the observation that ponatinib had comparable monomer and dimer RMSD in our 15-μs simulations suggests that more simulation time might be needed to reveal ponatinib's lower stability with monomeric BRAF (Supplementary Fig. S2A and S2B). A list of RAF inhibitors grouped according to their ex vivo potency in inhibiting monomeric or dimeric BRAF^V600E is shown in Fig. 2D.

The fact that certain RAF inhibitors preferentially bind the dimeric over the monomeric form of BRAF^V600E indicates that the two forms are able to adopt different conformations that are not captured by the available crystallographic data. Thus, to gain insight into these BRAF conformations, we used the formation of the MEK–BRAF–CRAF signaling complex during growth factor–induced MAPK activation to indirectly monitor the conformation of full-length BRAF as a function of its activation in the native state. As shown in Fig. 3A, MEK, BRAF, and CRAF interact weakly at steady state. EGF stimulation strongly promoted formation of the MEK–BRAF–CRAF complex within 5 minutes, followed by a gradual return to low basal levels of interaction after 30 to 60 minutes. The formation and subsequent disruption of the MEK–BRAF–CRAF signaling complex correlated well with activation and subsequent return to basal levels of MAPK signaling (Fig. 3B), suggesting that activated, dimeric BRAF interacts strongly with MEK, but inactive, monomeric BRAF adopts a conformation with low affinity for MEK (Fig. 3C). We further assessed the MEK–BRAF complex after inhibiting RAS activity in either RKO cells treated with the SHP2 inhibitor RMC-4550 (ref. 32; Fig. 3D) or in PC9 cells treated with the EGFR inhibitor GEF (Fig. 3E). In each case, drug treatment resulted in disruption of the MEK–BRAF complex, consistent with inactive, monomeric BRAF adopting a conformational state that interacts weakly with MEK.

We next asked whether forced dimerization of BRAF, CRAF, or BRAF^V600E using CID would increase their interaction with MEK. In fact, forced dimerization of either BRAF^V600E (Fig. 3F) or ectopically coexpressed BRAF and CRAF (Fig. 3G and H) markedly promoted their interaction with MEK. Finally, we examined the MEK–BRAF complex by coimmunoprecipitation experiments in additional settings in which BRAF^V600E exists either as a monomer or a dimer, including the two cell line pairs expressing full-length or splice variants of BRAF^V600E (SKMEL239 PAR and C3, M397-P, and M397-R) and a pair of Parental and a RAF inhibitor–resistant cell line (SK28-PAR and SK28-R) that we generated after continuous treatment of BRAF^V600E-expressing SKMEL28 melanoma cells to VEM. SK28-R expresses BRAF^V600E with duplicated kinase domain (Supplementary Fig. S3), similar to a RAF inhibitor–resistant cell line reported previously (33). In each case, we detected much higher levels of MEK interacting with the dimeric compared with the monomeric state of BRAF^V600E (Fig. 3I). Together, these data indicate that monomeric and dimeric full-length BRAF^V600E signal in different conformational states. Although both forms are highly catalytically active, monomeric full-length BRAF^V600E appears to be in a conformation that resembles catalytically inactive, monomeric BRAF^WT (i.e., with weak binding to MEK), whereas dimeric BRAF^V600E appears to be in a conformation that resembles active, dimeric BRAF^WT (i.e., with strong binding to MEK; Fig. 3J).

Despite recent progress in the structural characterization of RAF–MEK complexes (34, 35), how BRAF activation and dimerization affect its interaction with MEK remains elusive. In addition to our data presented here, it has been reported that the splice variant BRAF^V600E [Δex2–8], which contains only the kinase domain, interacts strongly with MEK, but this interaction is reduced when introduced mutations force it to adopt the monomeric state (i.e., R509H or S729A; ref. 36). These data suggest that the differential strength of the interaction between various BRAF states and MEK is at least in part determined by the BRAF kinase domain–MEK interactions. Comparisons between crystal structures of MEK bound to the active (PDB:4MNE) and inactive (PDB:6U2G) BRAF kinase domain did not reveal major differences at the enzyme–substrate interface, which is challenging to reconcile with the biochemistry data. Thus, to study the structural basis of the apparent biochemical difference between these states, we employed MD simulations.

We performed 25-μs simulations initiated from crystal structures of MEK bound to active and inactive BRAF, with three independent simulations for each condition. In all three simulations initiated from the active, αC-IN conformation, the N-terminal region of the BRAF αC-helix came into contact with the N-lobe of MEK. Specifically, several residues in the BRAF β3-αC loop interact with the P-loop on MEK. The interactions between MEK and the C-lobe of BRAF remained intact, as seen in crystal structures, during the simulations. As a consequence, the overall interfacial area between MEK and BRAF increases from the 1,312 Ų seen in the crystal structure to 1,630 Ų. In contrast, the inactive conformation of BRAF had a mobile αC-helix, which did not engage MEK, and the overall interfacial surface area was around 1,440 Ų at the end of the simulations. (Fig. 4A; Supplementary Videos S6 and S7).

We next sought to exploit the properties of RAF inhibitors to rationally design a combinatorial strategy for BRAF^V600E tumors. We showed previously that relief of negative feedback and induction of dimeric BRAF^V600E promotes resistance to current clinical monomer-selective RAF inhibitors, such that combining with RAF inhibitors that potently inhibit dimeric RAF would be effective in blocking resistance in this context (Fig. 4B). However, because BRAF^WT also signals as a dimer, such RAF inhibitors are predicted to have a lower therapeutic index when used as single agents (7). We previously showed that the combination of VEM with an αC-IN RAF inhibitor (TAK) was effective in melanoma and colorectal cancer BRAF^V600E, but that the addition of VEM only marginally improved the effectiveness of TAK alone (7). To identify αC-OUT RAF inhibitors that would better synergize with αC-IN RAF inhibitors, we assayed the effect on MAPK signaling of four αC-OUT RAF inhibitors [i.e., the three FDA-approved, VEM, DAB, and ENC, and the “Paradox Breaker” PLX7904, PB (37)] in combination with the αC-IN RAF inhibitor TAK (38). As shown in Fig. 4C, addition of VEM or PB did not enhance the inhibitory effect of TAK on MAPK signaling. By contrast, the combination of either DAB or ENC with TAK showed significantly greater inhibition of MAPK signaling compared with either inhibitor alone (Fig. 4C). It is reported that TAK has a slow off-rate (11). Therefore, it was possible that the off-rate of αC-OUT RAF inhibitors might affect their synergy with TAK. We thus carried out washout experiments and monitored recovery of MAPK signaling as an estimate of the relative off-rate of the αC-OUT RAF inhibitors in cells. pERK quickly rebounded after VEM or PB washout. However, suppression of MAPK signaling persisted for much longer after washout of DAB or ENC (Fig. 4D; Supplementary Fig. S4), suggesting a slow rate of dissociation of the compound from BRAF^V600E. Thus, the ability of DAB and ENC to synergize with the αC-IN RAF inhibitor appears to be associated with their much slower off-rate compared with VEM and PB, resulting in prolonged association with monomeric BRAF^V600E in cells.

We next biochemically characterized a panel of candidate MEK inhibitors to test in triple combination, including the three MEK inhibitors, TRAM, COB, and BIN, that have been approved in combination with RAF inhibitors for the treatment of BRAF^V600E melanomas. Previous work using a smaller number of MEK inhibitors has shown that certain MEK inhibitors can overcome feedback-induced MEK activation either by disrupting the RAF–MEK complex or by promoting the MEK–RAF complex while preventing phosphorylation by RAF (13, 15). To gain further insight into the biochemical properties of MEK inhibitors, we first compared potency and durability of pERK suppression by a panel of MEK inhibitors using the same concentrations for each compound. Under these conditions, we found TRAM to most potently and durably suppress pERK across the compounds tested (Fig. 5A).

We next tested these same MEK inhibitors for their effect on the MEK–RAF complex. Formation or disruption of the RAF–MEK complex is the result of both RAF activation due to relief of negative feedback upon MEK/ERK inhibition as well as a consequence of direct allosteric effects of inhibitor binding to MEK (39). To distinguish between the two effects, we first treated cells with an ERK inhibitor (SCH772984), which resulted in relief of negative feedback and formation of the RAF–MEK complex independently of inhibitor binding to MEK. Thus, we were able to determine the direct allosteric effect of inhibitor binding to MEK by subsequently treating with the MEK inhibitor and determining the disruption or formation of the RAF–MEK complex by coimmunoprecipitation. We found that most MEK inhibitors disrupted the MEK–RAF complex to various degrees, with TRAM showing the most potent disruption of the complex among the MEK inhibitors tested (Fig. 5B; Supplementary Fig. S5A). In contrast to most MEK inhibitors, CH5126766 (CH766) promoted formation of the MEK–RAF complex, as previously reported (14).

All MEK inhibitors tested occupy the same allosteric site in MEK and bind in a similar fashion. To explain the differences in the potency of disrupting the MEK–RAF complex across inhibitors, we asked whether the relative potency of each inhibitor in disrupting the complex correlates with its affinity for binding MEK in vitro and found that this was indeed the case (Fig. 5C; Supplementary Fig. S5B). In fact, after normalizing inhibitor concentrations based on the in vitro binding affinities, both COB and BIN potently disrupted the MEK–RAF complex, similar to TRAM (Fig. 5D), providing further evidence that despite the fact that most MEK inhibitors bind MEK in a similar structural fashion, the increased binding affinity of TRAM for MEK is the basis of the higher potency by which it disrupts the MEK–RAF complex.

To gain structural insight on the common features of most MEK inhibitors in disrupting the MEK–RAF complex, we overlaid the available atomic structures for MEK inhibitors used in this study. These MEK inhibitors bind the same allosteric site in MEK and interact with V211 and S212 (D+3 and D+4 positions from the DFG motif; ref. 13) of the activation loop. This interaction usually involves the fluoro or carbonyl moiety of the inhibitors to form hydrogen bonds with the backbone amide NH of V211 and S212, thereby stabilizing the local structure of the activation loop; (ref. 40; Fig. 5E). This conformation of the activation loop then positions the S212 hydroxyl side chain to form a hydrogen bond with the Q214 backbone amide, thus promoting the formation of the α-helix in the G213–M219 segment of the activation loop by “capping” the N-terminal end of the α-helix (Fig. 5E). Furthermore, the S212 hydroxyl side chain forms a hydrogen bond with E114 in the αC-helix, locking MEK in the unique αC-helix-OUT/DFG-IN (CODI) conformation (17). Docking of TRAM into atomic structures of MEK1 [PDB:3PP1 (41) and 3WIG (15)] revealed that the N-methylamide carbon of TRAM may have steric clashes with the side chains of L115 of αC-helix and V211 of the activation loop (Fig. 5E; Supplementary Fig. S5C). These steric clashes are unique to TRAM, as other MEK inhibitors do not interfere with L115 and V211. Thus, TRAM binds a local conformation of the αC-helix and activation loop different from other MEK–RAF “disruptor” inhibitors, which may account for its higher potency and durability in suppressing pERK in cells, compared with the other MEK inhibitors tested.

Although most MEK inhibitors induce an activation loop conformation that impedes the formation of the MEK–RAF complex, CH766 has a unique chemical moiety that alters the activation loop conformation and may be responsible for promoting the formation of an unproductive MEK–RAF complex. The atomic structure of the MEK–CH766 complex (PDB:3WIG; ref. 15) indicates that the core scaffold of CH766 interacts with V211 and S212 via its carbonyl moiety (Supplementary Fig. S5D) in a similar fashion to the “MEK–RAF disruptors.” However, unique to this atomic structure is the sulfamoyl moiety of CH766, which extends toward the A220–T226 segment of the activation loop and interacts with N221 in an alternative conformation not observed in other atomic structures of MEK1–inhibitor complexes (Fig. 5E; Supplementary Fig. S5D). This extended sulfamoyl NH interacts with the N221 side chain and stabilizes this unique local conformation of the activation loop that may promote the formation of the MEK–RAF complex, albeit in an alternative conformation that is not conducive for phosphorylation and results in “trapping” active RAF kinases in an unproductive MEK–RAF complex. To test this idea, we used CH4987655 (CH655), a MEK inhibitor that is structurally similar to CH766 but lacks the sulfamoyl moiety. Consistent with this model, treatment of cells with CH655 resulted in disruption (instead of promotion) of the RAF–MEK complex (Supplementary Fig. S5E), as well as failure to suppress MEK phosphorylation (ref. 15; Supplementary Fig. S5F). The relative biochemical properties of MEK inhibitors are summarized in Fig. 5F. As TRAM showed high potency in MEK binding, disruption of the RAF–MEK complex, and potent and durable suppression of pERK in cells, it was thus selected for subsequent studies in combination with RAF inhibitors.

To both overcome adaptive resistance and retain a high therapeutic index when targeting BRAF^V600E tumors, we used a strategy combining a RAF monomer–selective inhibitor with a RAF dimer–selective RAF inhibitor and a MEK inhibitor (Fig. 4B). As MAPK signaling in normal tissue is driven by RAF dimers, but in BRAF^V600 tumors from both RAF monomers and dimers, this strategy includes the RAF dimer inhibitor at relatively lower concentrations to avoid toxicities but sufficient to suppress the RAF dimers in the tumor (Fig. 6A). First, cell proliferation assays were carried out in the presence of various inhibitor combinations using BRAF^V600E colorectal cancer cell lines (WiDr and RKO) and one BRAF^V600E melanoma cell line (A2058) that are relatively insensitive to RAF monomer–selective inhibitors (7). We found that the triple combination of DAB + TRAM with a RAF dimer–selective inhibitor (REG, LXH254, or RAF709) resulted in all cases in more potent suppression of colony formation, compared with TRAM alone, or combined DAB with TRAM (Fig. 6B). To obtain an estimate of the potential therapeutic index of the triple combination, we used cells WT for both BRAF and RAS (BRAF^WT) as a surrogate for normal cells and compared the effect of the triple combination on MAPK signaling in BRAF^WT versus BRAF^V600E cells. The triple combination REG + DAB + TRAM potently suppressed MAPK signaling in the BRAF^V600E cell lines compared with the BRAF^WT cell lines, and in BRAF^WT cells, the presence of DAB reduced the inhibitory effect of REG + TRAM on MAPK signaling, suggesting a broad therapeutic window of the triple combination (Fig. 6C; Supplementary Fig. S6A and S6B). Furthermore, potent suppression of the MAPK rebound by REG was not an indirect effect of receptor tyrosine kinase suppression, as both total and phosphorylated levels of EGFR were not affected by the presence of REG (Supplementary Fig. S6C).

Finally, as therapies targeting components of MAPK signaling have been shown to affect tumor response to immunotherapy (42), we examined the effect of the triple combination on T-cell signaling and function. Markers of T-cell receptor (TCR) signaling, such as phosphorylated ZAP70 (pZAP70), pPLC1γ, pAKT, and pERK, were at similar levels when primary mouse T cells were treated with DAB + TRAM, REG alone, or REG + DAB + TRAM, whereas the Src and multikinase inhibitor dasatinib (DAS) potently suppressed TCR signaling (Fig. 6D; Supplementary Fig. S6D). Furthermore, treatment of activated mouse primary T cells with DAB + TRAM or REG + DAB + TRAM resulted in similar levels of secretion of IL2 or IFNγ, whereas DAS suppressed both IL2 and IFNγ secretion (Fig. 6E). To assess the effect of the triple combination on tumor killing by T cells, we used the Just eGFP Death Inducing (JEDI) T cells (43, 44). JEDI T cells recognize the immunodominant epitope of GFP (GFP200–208) presented in MHC class I. JEDI cells were cocultured with GFP- or mCherry-expressing tumor cells at a ratio of 1:1, and killing was assessed by flow cytometry at day 3 after coculture measuring the relative percentage of GFP⁺ and mCherry⁺ cells. Treatment with REG + DAB + TRAM resulted in tumor cell killing at levels comparable to or higher than DAB + TRAM (Supplementary Fig. S6E). Together, these data indicate that the triple combination will potently and durably suppress MAPK signaling in the tumor without negatively affecting tumor cell killing by T cells.

To assess the effectiveness of the combinatorial strategy in vivo, we first treated mice carrying xenograft BRAF^V600E colorectal cancer tumor models (WiDr or RKO) with REG + DAB + TRAM and compared the effect with the double combinations REG + TRAM or DAB + TRAM. In both xenografts, either control or DAB + TRAM treatment had only minimal effect. However, REG + DAB + TRAM or REG + TRAM combinations potently suppressed MAPK and downstream signaling as well as tumor growth (Fig. 7A; Supplementary Fig. S7A). Similar results were obtained using LXH254 instead of REG. As shown in Fig. 7B, the triple LXH254 + DAB + TRAM combination potently suppressed growth of WiDr xenografts in vivo. We next assessed the in vivo effectiveness of the triple combination using two patient-derived xenograft (PDX) BRAF^V600E melanoma models (WM4262 and WM4398) that are derived from tumors resistant to clinical RAF and MEK inhibitors (45). In both PDXs, control treatment and DAB + TRAM had minimal effect, while treatment with REG + DAB + TRAM potently suppressed tumor growth (Fig. 7C).

Limitations related to therapeutic index and potential added toxicities are major considerations when designing drug combinations. When treating BRAF-mutant tumors, the therapeutic index is expected to derive from the fact that two RAF inhibitors (i.e., a RAF monomer–selective and RAF dimer–selective) synergize in the tumor but exert opposite effects on MAPK signaling in BRAF^WT cells. Thus, in normal cells, combining a RAF dimer–selective with a RAF monomer–selective inhibitor is expected to maintain MAPK signaling close to basal levels, consistent with our in vitro data shown in Fig. 6B and Supplementary Fig. S6B. Our in vivo observations were consistent with this concept. A gradual weight loss was observed in mice on the double combination REG + TRAM after the third week of treatment, suggesting accumulating toxicities caused by this treatment. Strikingly, mice receiving the triple combination REG + DAB + TRAM or LXH254 + DAB + TRAM showed no evidence of weight loss or other apparent toxicities (Fig. 7D; Supplementary Fig. S7B and S7C). Together, these data show that by rationally designing combinations, addition of a compound targeting a conformational state of RAF could rescue from toxicities caused by another conformation-selective RAF inhibitor and thus increase the therapeutic index.

To assess this therapeutic strategy in the clinical setting, the combination of REG + DAB + TRAM was tested in a 53-year-old man with stage IV sigmoid colon adenocarcinoma and peritoneal carcinomatosis whose disease had progressed on standard cytotoxic chemotherapies. Molecular analysis demonstrated the presence of a BRAF^V600E mutation; the tumor was also microsatellite stable, RAS^WT, and HER2 negative. The patient's serum carcinoembryonic antigen (CEA) level at diagnosis was 18.3 ng/mL but increased to 39.1 ng/mL when he began therapy with FOLFOX. Due to a lack of serologic response and radiographic evidence of disease progression after 2.5 months of therapy, he was transitioned to FOLFIRINOX for 4 months. Although his disease appeared stable on imaging and he maintained a good functional status, his appetite remained poor, and he continued to report lower abdominal pain associated with an abrupt increase in CEA level to 115.2 ng/mL.

Given our preclinical data and the poor prognosis associated with BRAF^V600E, which compels the earlier use of combination targeted therapy, compassionate use of off-label REG + DAB + TRAM was initiated, including DAB and TRAM at 150 mg twice a day orally and 2 mg every day orally, respectively, as used in melanoma (46). Given that REG has never been combined with DAB + TRAM, and based on the ReDOS trial (47), a starting REG dose of 40 mg every day orally for 1 week was selected, with the plan to increase to 80 mg every day during week 2 and 120 mg every day during week 3 of therapy. Serum CEA at the start of treatment, approximately 4 weeks after the patient's last dose of FOLFIRINOX, was 129.6 ng/mL (Fig. 7E). The first month on REG + DAB + TRAM was complicated by grade 2 to 3 cough, gingival bleeding and epistaxis, pyrexia, rigors, and pancytopenia. The patient was hospitalized twice: first, for thrombocytopenia and Escherichia coli bacteremia and, second, for pyrexia and rigors without a documented infection. Medication reconciliation revealed underdosing of DAB and a doubling of the REG starting dose. Notably, his CEA level decreased significantly to 9.8 ng/mL, and repeat imaging showed stable disease. However, during the 4-week interruption in therapy for toxicities, the patient's CEA increased to 24.7 ng/mL. On February 13, 2020, he resumed modified-dose TRAM 1.5 mg every day orally, DAB 100 mg twice a day orally, and a flat dose of REG 40 mg every day orally, which had been well tolerated without further hospitalizations, and the patient's CEA dramatically decreased, appetite improved, and lower abdominal pain decreased. The triple-combination treatment continued for a total of almost 8 months, until August 17, 2020, when an increase in CEA was observed along with mild disease progression (Fig. 7E). Moreover, after experiencing weight loss on the prior regimen, the patient's weight has remained stable during the triple-combination regimen.

Several lines of evidence indicate that adaptive drug resistance is frequently the result of incomplete inhibition of oncogenic signaling, and to effectively overcome it, combinatorial targeting of multiple nodes of the pathway will be necessary (9, 10, 39, 48–53). However, how to rationally design drug combinations with high effectiveness and minimal toxicities remains a major challenge. Here we identify and characterize a novel class of RAF inhibitors that selectively inhibits dimeric over monomeric RAF. The class includes a number of clinically relevant compounds, such as the FDA-approved multikinase inhibitor REG, as well as LHX254 and belvarafenib (GDC-5773), which are currently in clinical trials. Although this class of compounds shows very similar structural and crystallographic properties to compounds equipotent for monomeric and dimeric RAF, MD simulations provide a structural explanation for their dimer selectivity: higher inhibitor stability in the dimer specificity pocket due to stabilization of the αC-helix upon RAF dimerization. The selectivity of these compounds revealed a previously unknown difference in the conformation of the active site between dimeric and monomeric BRAF^V600E and their affinity for MEK, thus uncovering a hitherto unappreciated complexity in the regulation of the BRAF^V600E oncoprotein. A previous report indicated that BRAF^V600E shows weaker interaction with MEK compared with RAS-activated WT BRAF, a phenotype attributed to lower affinity of highly phosphorylated MEK for BRAF (54). Another report showed that the splice variant of BRAF^V600E interacts more strongly with MEK compared with full-length BRAF^V600E (36). Two recent reports presented structural data using cryo–electron microscopy on BRAF (34, 35). Park and colleagues (35) studied inactive and monomeric BRAF in complex with MEK. However, based on our findings, in cells, BRAF in the inactive, monomeric conformation interacts much more weakly with MEK compared with dimeric BRAF. The discrepancies between structural and biochemical findings across different studies highlight the complexity of BRAF signaling and the need for further biochemical and structural analysis to conclusively elucidate the regulation of WT and mutant BRAF in their native states in different cellular contexts.

After characterizing a number of RAF and MEK inhibitors, we assessed the therapeutic strategy of the combination of a RAF monomer–selective, a RAF dimer–selective, and a potent MEK–RAF “disruptor” MEK inhibitor to overcome adaptive resistance of BRAF^V600E tumors. The triple combination was highly effective and well tolerated in multiple BRAF^V600E tumor models in vivo as well as in a patient with advanced colorectal cancer, suggesting that this may be a powerful therapeutic strategy for patients with BRAF^V600E tumors. In the phase III BEACON trial, the combination of a RAF inhibitor (ENC) and an anti-EGFR antibody (cetuximab) in patients with metastatic BRAF^V600E colorectal cancer tumors showed a median progression-free survival (PFS) of 4.2 months and median overall survival (OS) of 8.4 months, which represented an improvement over standard chemotherapy (55), resulting in approval of the combination for this indication by the FDA. Although case reports must be interpreted with caution, the response of our patient to REG + DAB + TRAM for almost 8 months compares favorably to prior studies of DAB + TRAM (median PFS 3.5 months; ref. 56) and REG monotherapy (median PFS 1.9 months; ref. 57) in BRAF^V600E colorectal cancer. The combination also compares favorably to the BEACON trial results and to other combinations tested in this population such as irinotecan + cetuximab + VEM in the SWOG 1406 study (median PFS 4.2 months; ref. 58) and DAB + TRAM + panitumumab (median PFS 4.2 months, median OS 9.1 months, median response duration 7.6 months; ref. 59). This observation and the clear clinical benefit experienced by the patient provide anecdotal support for the clinical translation of the triplet regimen as a potentially active and tolerable option in patients with BRAF^V600E mutant colorectal cancer. The close temporal association between the initiation, interruption, and reinitiation of therapy with changes in serum CEA levels and the subjective improvement experienced by the patient compellingly indicate the antineoplastic effect of this regimen, providing a basis for further evaluating this combination in a clinical trial.

The initial enthusiasm based on the objective clinical effectiveness of RAF and MEK inhibitors in BRAF^V600E patients was subsequently tempered by the realization that for most of these patients, the antitumor effect of these drugs is temporary, frequently limited by various mechanisms of adaptive and acquired resistance. The discovery of a new class of RAF inhibitors described here paves the way for the rational design of next-generation combinatorial pharmacologic strategies with high efficacy and increased therapeutic index for therapy of patients with BRAF^V600E tumors.

Cells were washed with PBS and lysed on ice for 5 minutes in NP40 buffer (50 mmol/L Tris pH 7.5, 1% NP40, 150 mmol/L NaCl, 10% Glycerol 1 mmol/L EDTA) supplemented with protease and phosphatase inhibitors (Roche). Lysates were centrifuged at 15,000 rpm for 10 minutes, and the protein concentration was quantified using BCA (Pierce). Proteins were separated by NuPAGE and 4% to 12% Bis Tris Gel (Novex), and they were immunoblotted and transferred to nitrocellulose membranes (GE Healthcare) according to standard protocols. Membranes were immunoblotted overnight with antibodies against pMEK1/2^Ser217/221, MEK1 (61B12), pERK1/2^{Thr202/Tyr204} (D13.14.4E), ERK (137F5), pEGFR^Tyr1068 (D7A5), pBRAF^Ser445, pZAP70^Tyr319, pPLCγ1^Tyr783, pAKT^Ser473 (D9E), pSTAT1^Tyr701 (58D6), EGFR (D38B1), and β-actin (13E5) from Cell Signaling; BRAF from Santa Cruz Biotechnology; BRAF^V600E from NewEast Biosciences; hemagglutin (HA) and DUSP6 from Abcam; CRAF from BD Transduction Laboratories; and pCRAF^Ser338 from Millipore. The next day, membranes were probed with anti-rabbit IgG or anti-mouse IgG secondary antibody (Cell Signaling), and chemiluminescent signals were detected on X-ray films. All the Western blot experiments were performed twice, unless stated otherwise, and representative results from each experiment are shown.

For immunoprecipitations, lysates were incubated with the indicated antibodies overnight at 4°C, followed by protein G agarose (Life Technologies) for 1 hour at 4°C. Samples were washed three times with lysis buffer, and sample buffer was added for subsequent immunoblot analysis.

RKO, HT29, WiDr, A375, A2058, HCT116, and Calu-6 cell lines were purchased from ATCC. 293H cells were purchased from Life Technologies. HeLa cells were kindly provided by Ramon Parsons (Icahn School of Medicine at Mount Sinai), and 3T3 cells were developed by Stuart Aaronson. MEF, COS7, PC9, 239/C3, M397-P/-R, WM1382, and SKMEL28(SK28)/SKMEL28(SK28)-R cell lines have been described previously (7). All the cells used in the study were maintained in a humidified incubator at 37°C with 5% CO₂ and cultured in RPMI 1640 or DMEM supplemented with 10% FBS, 2 mmol/L glutamine, and 100 IU/mL penicillin and streptomycin, and were passaged from three to five times. Cell lines were authenticated by LabCorp using short tandem repeat DNA profiling and regularly tested negatively for Mycoplasma using the Venor GeM Mycoplasma Detection Kit (Sigma).

CH655 and RMC-4550 were obtained from Medchem Express. PB was kindly provided by Plexxikon. REG was obtained from Selleckchem or Medchem Express. All other compounds were obtained from Selleckchem. Compounds were dissolved in DMSO to yield 10 mmol/L stock. The ligand for chemically induced dimerization AP (Clontech) was dissolved in ethanol.

Cells were treated with 2 μmol/L or 4 μmol/L VEM or REG for 2 hours. Cells were resuspended in PBS containing protease inhibitor cocktail (Roche), and the cell suspensions were aliquoted into 100-μL PCR tubes. Samples were then heated individually at the indicated different temperature endpoints for 3 minutes using the Veriti 96-well thermal cycler (ThermoFisher) and allowed to equilibrate to room temperature. The cell suspensions were freeze-thawed three times using liquid nitrogen and a thermal cycler at 25°C. The soluble fractions (lysates) were separated from the cell debris by centrifugation at 20,000 × g for 20 minutes at 4°C. Supernatants were collected and analyzed by SDS-PAGE, followed by Western blot analysis.

Cells were plated in 6-well plates at a density of 1 to 10 × 10³ cells per well. The next day, cells were treated with inhibitors as indicated in regular growth media for 10 to 14 days. Growth media with or without inhibitors were replaced every 3 days. Cells were fixed with 4% paraformaldehyde for 5 minutes and then stained with 0.5% crystal violet for 30 minutes. Cells were destained with tap water and air-dried.

iDimerize inducible system (Clontech) was used for chemical induced dimerization. BRAF^V600E, BRAF, or CRAF was cloned at the N-terminus of the DmrB domains of the pHom–Mem1 vector using EcoRI and XbaI restriction sites. The vector was designed to express an HA epitope tag at the C-terminus of the created fusion proteins. Transfections were carried out using Lipofectamine 2000 (Life Technologies).

The dissociation constant (Kd) values for the binding between TRAM, COB, BIN, or CH766 and MEK1 were determined using the KinomeScan binding assay (60). Binding studies were performed by Eurofins DiscoverX.

CD4⁺ and CD8⁺ T cells were isolated from the spleens of C57BL/6 mice (Envigo Laboratories) through negative selection using the EasySep mouse T-cell isolation kit (StemCell Technologies). The isolated T cells were activated for 48 hours using magnetic beads coupled with antibodies against CD3 and CD28 (Dynabeads Mouse T-Activator CD3/CD28; ThermoFisher) and 20 ng/mL recombinant IL2 (Peprotech) according to manufacturer instructions in RPMI supplemented with 10% FBS, 100 U/mL penicillin/streptomycin, 2 mmol/L L-glutamine, 1% nonessential amino acids, 1 mmol/L sodium pyruvate, 55 μmol/L 2-mercaptoethanol, and 20 mmol/L HEPES.

Activated murine primary T cells were plated in 96-well plates at a density of 1 × 10⁵ cells per well and treated with different inhibitors or inhibitor combinations for 24 hours. Supernatants were collected, and the concentrations of IFNγ and IL2 were determined using the ELISA MAX Standard Set Mouse IFNγ (BioLegend) and the uncoated mouse IL2 ELISA (Thermo Fisher), respectively, according to manufacturer instructions.

CD8⁺ T cells were isolated from spleens of JEDI mice (43). Splenic cells suspensions were obtained by mechanical disruption and filtering through a 70-μmol/L cell strainer. Red blood cells were lysed using an RBC buffer (eBioscience), and CD8⁺ T cells were negatively selected using the EasySep mouse CD8⁺ T cell isolation kit from STEMCELL Technologies following manufacturer instructions. Cells were activated for 48 hours with Dynabeads Mouse T-Activator CD3/CD28 (Thermo Fisher) and 20 ng/mL mouse recombinant IL2 (Peprotech) in RPMI with 10% FBS, 100 U/mL penicillin/streptomycin, 2 mmol/L L-glutamine, 1% nonessential amino acids, 1 mmol/L sodium pyruvate, 55 μmol/L 2-mercaptoethanol, and 20 mmol/L HEPES. Target or bystander cells were generated by transducing MEL (mouse-derived erythroleukemia cell line) cells with a lentiviral vector expressing GFP or mCherry and then sorted to purity. A 50%/50% mix of GFP⁺ (10⁴ target cells) and mCherry⁺ (10⁴ bystander cells) MEL cells were plated in a 96-well plate (10⁴ target cells per well), and activated T cells were added at different ratios along with different inhibitors or inhibitor combinations. Killing was assessed by flow cytometry at day 3 after coculture, measuring the relative percentage of GFP⁺ and mCherry⁺ cells.

Docking of TRAM was performed with the Glide program of the Schrödinger suite (Schrödinger release 2019–2: LigPrep, Glide, Protein Preparation Wizard; Schrödinger, LLC).

TRAM was prepared and tautomerized at pH 7.2 by LigPrep and MEK kinase domain structures (PDB:3PP1 and 3WIG) were prepared by Protein Preparation Wizard. Ligand and protein were parameterized with OPLS3 force field (61). Protein grid generation has the following settings: (i) aromatic CH hydrogens and halogen atoms were treated as hydrogen-bond donors and acceptors, respectively, and (ii) van der Waals radius was softened (scaled to 0.8) for atoms with partial charge exceeding 0.25e. The standard Precision mode of Glide was used for molecular docking.

Proteins and ions were parameterized with the Amber99SB*–ILDN force field (62, 63), and ligands were parameterized using GAFF 2.0 (64). The system was solvated with water parameterized with the TIP3P force field (65) and neutralized with a 100-mM NaCl buffer. Monomeric BRAF systems contained ∼55,000 atoms in a cubic box of length 85 Å, and dimeric systems contained ∼110,000 atoms with a cubic box of length 103 Å. BRAF–MEK systems contained ∼121,000 atoms with a cubic box of length 107 Å.

Systems were each equilibrated on GPU Desmond using a mixed NVT/NPT schedule (66) and in https://www.deshawresearch.com/publications/Desmond-GPU%20Performance%20as%20of%20October%202015.pdf. All production simulations were run on Anton, a specialized machine for MD simulations (67). Simulations were performed in the NPT ensemble at 310 K using the Martyna–Tobias–Klein barostat (68). The simulation time step was 2.5 fs, using a modified r-RESPA integrator (69, 70) and evaluating long-range electrostatics every three time steps. Electrostatics forces were calculated using the u-series method (71). A 9-Å cutoff was applied for the van der Waals calculations.

BRAF crystal structures from the Protein Data Bank were prepared for simulation using the Protein Preparation Wizard in Schrödinger Maestro (Schrödinger release 2019–2: Maestro; Schrödinger, LLC). Unless otherwise noted, missing loops and termini were capped with ACE/NME capping groups.

Both protomers of a BRAF dimer contain a ligand molecule. The monomeric system was generated by deleting one copy of the dimeric system and then solvating as previously described. Ligands were prepared and protonated using LigPrep, with EPIK at pH 7.0. Simulations of TAK, AZ628, LY3009120, RAF709, and BGB283 were initiated from their respective crystal structure poses. Simulations of ponatinib, LHX254, SOR, REG, and belvarafenib were initiated from docked poses, generated using Standard Precision Glide with default receptor grid generation parameters. For each ligand, three simulations of 15 μs were run for both the dimer and the monomer. Ligand placement methodology and starting crystal structures are summarized in Supplementary Table S1.

BRAF–MEK structures were prepared from PDB IDs 4MNE (54) and 6U2G (72) using the Protein Preparation Wizard. In the 6U2G structure, the unresolved activation loop of BRAF, from residues 602 to 610, was built using Schrödinger Prime (73, 74). Both BRAF–MEK constructs were simulated for 15 μs.

All simulations were visually inspected using the in-house visualization software Firefly. Ligands remained bound to the kinase throughout the simulations. For a given ligand, the chemical structure that interacted with the specificity pocket (as defined by residues 504, 505, 513, 567, and 572) exhibited larger dynamic fluctuation than the regions that interacted with the classic ATP binding pocket. To understand the differences between monomeric and dimeric BRAF ligand simulations, we measured the RMSD of the portion of the ligand that binds in the specificity pocket. These portions were manually selected and correspond to the circled heavy atoms in Supplementary Table S1. The system was aligned on all ligand-heavy atoms. For dimeric simulations, only the ligand bound to chain A was considered. RMSD data across the replicates were aggregated to produce the histograms in Fig. 2A and Supplementary Fig. S2A. The RMSD of the αC-helix was quantified in a similar manner. The system was aligned on the alpha carbons of chain A residues 490 to 510, and the RMSD of this same selection was computed. These RMSD values were aggregated in the histograms in Supplementary Fig. S2B.

As an additional metric to evaluate the αC-helix dynamics, we report in Supplementary Table S2 the stability of the E501–K483 salt bridge (calculated by measuring the fraction of the trajectory in which the distance between the two residues was smaller than 3.5 Å).

For the BRAF–MEK dimer simulations, protein–protein contact areas were computed on the final frame of the simulation using the PDBe PISA server at the European Bioinformatics Institute (75). Reported interface areas are the per-protomer area, defined as the solvent-accessible surface area of the complex subtracted from the sum of the protomer areas and divided by 2.

All animals were examined prior to the initiation of studies to ensure that they were healthy and acclimated to the laboratory environment. Five- to 7-week-old female athymic Nude–Foxn1^nu (Envigo Laboratories) mice were used for animal experiments. All mouse experiments were approved by the Icahn School of Medicine at Mount Sinai Animal Care and Use Committee (protocol no. IACUC-2016–0066). Mice were maintained under specific pathogen–free conditions, and food and water were provided ad libitum.

RKO or WiDr cells were harvested on the day of use and injected subcutaneously in the left flank per mouse (10 × 10⁶ per injection). After inoculation, mice were monitored daily and weighed every 3 days, and caliper measurements were begun when tumors became visible. Tumor volume was calculated using the following formula: tumor volume = (D × d2)/2, in which D and d refer to the long and short tumor diameter, respectively. When tumors reached a size of 100 to 150 mm³, mice were randomized into seven groups (n = 7/group) treated each with vehicle (5% DMSO, 0.5% hydroxypropyl methyl cellulose, 0.2% Tween 80); REG (30 mg/kg); LXH254 (30 mg/kg); DAB (30 mg/kg) and TRAM (0.25 mg/kg); REG (30 mg/kg) and TRAM (0.25 mg/kg); the triple combination of REG (30 mg/kg), DAB (30 mg/kg), and TRAM (0.25 mg/kg); or the triple combination of LXH254 (30 mg/kg), DAB (30 mg/kg), and TRAM (0.25 mg/kg) orally once a day based on mean group body weight. The formulation for LXH254 was prepared after dilution with DI water of the microemulsion stock vehicle MEPC4 (45% cremophor RH40, 27% PEG400, 18% corn oil glycerides, and 10% ethanol). The endpoint of the experiment for survival studies was considered a tumor volume of 1,000 mm³ as per our approved protocol. Once the mice were sacrificed, tumors were used for further analysis.

NSG (NOD.Cg-PrkdcscidIl2rgtm1Wjl/SzJ) mice models were injected with single-cell suspensions of the indicated PDX melanoma cells (WM4262 and WM4398). When the tumors were palpable, mice were randomized into three groups (n = 7/group): vehicle control, DAB (30 mg/kg) and TRAM (0.3 mg/kg), or the triple combination of REG (30 mg/kg), DAB (30 mg/kg), and TRAM (0.3 mg/kg). Mice were fed chow containing the DAB and TRAM daily. Mice were treated with REG 5 days on, 2 days off for the duration of the therapy trial. All the animal experiments regarding the PDX models were approved by the Wistar Institute Institutional Animal Care and Use Committee (76).

Data are presented as mean with SEM, as indicated in the figure legends. Statistical comparisons were performed using two-sided unpaired Student t test, using GraphPad Prism 9 (GraphPad Software), unless otherwise specified (*, P < 0.05; **, P < 0.01; ***, P < 0.001).

B.D. Brown reports a patent for EP3043641B1 issued. A. Schlessinger is the cofounder of AIchemy. All work on this study performed by D.E. Shaw, Q. Wang, and M.R. Tucker was conducted within and funded internally by D. E. Shaw Research LLC, of which D.E. Shaw is the sole beneficial owner and Chief Scientist and with which Q. Wang and M.R. Tucker are affiliated. In the interest of maximal disclosure, we note that D. E. Shaw Research LLC has scientific and commercial collaborations unrelated to this study with various outside entities in the biopharmaceutical sector, including Relay Therapeutics, Inc. and Schrödinger, Inc.; none of these outside entities were involved with, or exerted any influence over, the work for this study done by D.E. Shaw, Q. Wang, or M.R. Tucker. No disclosures were reported by the other authors.

C. Adamopoulos: Data curation, investigation, formal analysis, writing-original draft. T.A. Ahmed: Data curation, investigation, formal analysis, writing-original draft. M.R. Tucker: Data curation, investigation, formal analysis, writing-original draft. P.M.U. Ung: Data curation, formal analysis, investigation. M. Xiao: Data curation, formal analysis, investigation. Z. Karoulia: Data curation, investigation. A. Amabile: Data curation, formal analysis, investigation. X. Wu: Data curation, investigation. S.A. Aaronson: Supervision, writing–review and editing. C. Ang: Investigation, writing–review and editing. V.W. Rebecca: Data curation, formal analysis, supervision. B.D. Brown: Supervision. A. Schlessinger: Formal analysis, supervision. M. Herlyn: Supervision. Q. Wang: Formal analysis, supervision, investigation, writing-original draft. D.E. Shaw: Conceptualization, supervision, writing-review and editing. P.I. Poulikakos: Conceptualization, supervision, funding acquisition, writing–original draft, writing–review and editing.

The authors thank Paul Maragakis for helpful discussions. P.I. Poulikakos is supported by Tisch Cancer Institute (TCI) developmental awards, the NIH/NCI (R01CA204314, R01 CA240362, and R01CA238229), the Irma T. Hirschl Trust, the Manhasset Women's Coalition Against Breast Cancer, the Breast Cancer Alliance, the Melanoma Research Foundation, and the Melanoma Research Alliance. M. Herlyn is supported by NIH grants RO1 CA238237, U54 CA224070, and PO1 CA114046 and the Dr. Miriam and Sheldon G. Adelson Medical Research Foundation. B.D. Brown was supported by funding from the Cancer Research Institute and NIH R01AT011326 and R33CA223947. S.A. Aaronson is supported by the Breast Cancer Research Foundation. T.A. Ahmed is supported by grant T32CA078207. Z. Karoulia received the 2017 Robin Chemers Neustein Postdoctoral Fellowship.

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.