Distinct Transcriptional Programming Drive Response to MAPK Inhibition in BRAF^V600-Mutant Melanoma Patient-Derived Xenografts Free

Approximately 50% of cutaneous melanoma harbor oncogenic mutations of BRAF. The majority of BRAF mutations occur in the V600 position, resulting in constitutive activation of the MAPK signaling pathway (1, 2). The inhibition of BRAF, either alone or in combination with MEK, provides remarkable clinical benefits for patients with advanced BRAF^V600-mutant melanoma (3–6). Nevertheless, despite the robust initial response, most patients eventually develop resistance. Acquired resistance to MAPK inhibition involves heterogeneous genetic and nongenetic mechanisms that activate MAPK or other compensatory signaling pathways (7–9). As such, there is a need for patient stratification and treatment strategies in order to improve response and delay the onset of disease progression.

A subset of BRAF^V600-mutant melanoma patients progress rapidly upon treatment using targeted therapy (3–6). In contrast, approximately 20% of patients benefit from durable response to BRAF and MEK inhibition, as measured by progression-free survival at 3 years (10, 11). The disparity between these 2 patient groups is not well understood. Clinical factors related to disease burden, including lactate dehydrogenase levels and the number of metastases, are important prognostic markers (10, 11). At the molecular level, melanoma cell line sensitivity to BRAF-targeted therapy is linked to the melanoma cell state (12, 13). Gene-expression profiling revealed that melanoma falls into two distinct transcriptional states irrespective of their mutation status, characterized by proliferative or invasive features and MITF expression levels (14–16). The invasive, or MITF-low and AXL-high phenotype, confers intrinsic resistance to MAPK inhibition in BRAF^V600-mutant melanoma in vitro (12, 13). However, both high and low MITF expression levels are implicated in therapeutic resistance against MAPK inhibition in melanoma (17). It remains to be determined whether the melanoma cell state predicts response to BRAF inhibitor–based therapy in the clinical setting.

In this study, we sought to elucidate mechanisms governing differential sensitivity to BRAF and MEK inhibition, using our internally established patient-derived xenograft (PDX) collection (18). Work by our group and others demonstrated the prognostic value of PDXs in modeling clinical response to targeted therapies (18–20). In comparison with other preclinical experimental systems, PDX models conserve tumor architecture and heterogeneity, and faithfully recapitulate patient response to chemo- and targeted therapies (18, 19, 21–23). Moreover, because obtaining high-quality research biopsies from the clinic is challenging, PDXs enable profiling of treatment and time matched tumors with statistically powered replicates. Here, we characterized on-treatment PDX tumor samples to compare the pharmacodynamic effects of BRAF and MEK inhibitors in models that achieved complete regression versus partial response and stable disease. Transcriptome profiling revealed that in vivo, a MITF-high “epithelial-like” phenotype is associated with adaptive and intrinsic resistance, whereas a MAPK-high “mesenchymal-like” phenotype is linked to sensitivity to BRAF and MEK inhibitor treatment. This is different from findings using in vitro cancer cell lines (12, 13, 17) and suggests an in vitro–in vivo discrepancy that warrants additional investigation. In less sensitive models, the lack of tumor regression in response to treatment was model dependent and mediated by either MAPK-dependent or -independent mechanisms. Our results suggested that response of BRAF^V600-mutant melanoma xenografts to BRAF and MEK inhibitors is heterogeneous, and precision medicine–based combination strategies are required to target this aggressive cancer. Importantly, this study provided insights into determinants of therapeutic response in BRAF^V600 melanoma in vivo.

Mice were maintained and handled in accordance with the Novartis Institutes for BioMedical Research Animal Care and Use Committee protocols and regulations. PDX models were procured and established as previously described (18). Surgical tumor tissues were obtained from treatment-naïve cancer patients, all of whom provided informed written consent for the samples procured by Novartis Inc.

PDX fragments were implanted subcutaneously into the right flanks of female nude mice (Charles River) using a trocar. Once the tumor volumes reached 300 to 500 mm³ in size, mice were randomized into respective groups. For tissue collection, mice were dosed twice daily by oral gavage with encorafenib and binimetinib at 20 and 3.5 mg per kg, respectively. These doses were selected as they achieved exposures in mice comparable with that in humans (4). For efficacy studies, encorafenib and binimetinib were administered via diet supplementation (200 mg encorafenib and 35 mg binimetinib per kg of chow; Bio-Serv). The supplemented diet was designed to achieve similar drug exposure to twice-daily oral gavage. The ERK inhibitor VX-11e was administered orally at 50 mg per kg daily. Body weight and tumor size were monitored twice per week. Tumor volume was measured using a caliper and calculated as (length × width × width)/2. Best average response and response calls were determined as previously described (18). All studies were performed using PDXs between passages 4 and 9.

Total RNA were extracted from snap-frozen tumor fragments, using the RNeasy kit with the QIAcube (QIAGEN) automated sample preparation platform. The RNA concentration and integrity were measured using the RNA 6000 nano kit (Agilent Technologies) on the Agilent 2100 BioAnalyzer.

Two hundred nanograms of high-purity RNA was used as input to generate sample libraries using the TruSeq Stranded mRNA library prep kit (Illumina), following the manufacturer's instructions on the Hamilton STAR robotics platform. The PCR-amplified RNA-seq library products were quantified using the Fragment Analyzer Standard Sensitivity NGS Fragment Analysis Kit (Advanced Analytical Technologies). Samples were diluted to 10 nmol/L in Elution Buffer (QIAGEN), denatured, and loaded at a range of 2.5 to 4.0 pmol/L on an Illumina cBOT using the HiSeq 4000 PE Cluster Kit (Illumina). RNA-seq libraries were sequenced on a HiSeq 4000 at 75 base pair paired-end with 8 base pair dual indexes using the HiSeq 4000 SBS Kit, 150 cycles (Illumina). The sequence intensity files were generated on instrument using the Illumina Real-Time Analysis software and resulting intensity files were demultiplexed with the bcl2fastq2.

Gene-level read counts were normalized and voom transformated (24) for differential expression using the limma framework (25), and P values were corrected for multiple testing using the Benjamini–Hochberg method. Genes were considered differentially expressed if the log₂-fold change > 1.5 in absolute value and the adjusted P value < 0.05. Log-transformed and TMM-normalized (26) expression values were used for clustering analysis with NbClust (27). K = 5 was selected as the most informative partition. A hypergeometric distribution function was used to determine enrichment P values for overlap of each cluster membership with gene sets from mSigDB (28), in addition to transcription factor and canonical gene sets exported from MetaCore (Clarivate Analytics). Gene set enrichment P values were adjusted using the Benjamini–Hochberg method. Mutation and copy-number calls were made as described in Gao et al. (18). The RNA-seq data have been submitted to the Sequence Read Archive database; submission number PRJNA562481.

Normalized and background corrected microarray data were downloaded from Gene-Expression Omnibus public database, accession number GSE99898 (29). The data set was derived from 17 patients treated with dabrafenib, vemurafenib, or dabrafenib and trametinib. Seventeen untreated, 8 on-treatment, and 13 resistant samples were collected in total. Patient response data were obtained from Table 1 in the reference publication (29). Gene signature scores were calculated as sum of the normalized and z-transformed gene-expression values for all genes of interest.

Snap-frozen tumor fragments were suspended in RIPA lysis buffer with protease (Roche) and phosphatase (Sigma) inhibitor cocktails, and mechanically homogenized with tungsten beads using the TissueLyser (QIAGEN). Protein concentrations were quantified using the Pierce BCA protein assay kit (Thermo Fisher Scientific). Western blotting was performed using standard procedures. Antibodies against MEK1/2, pMEK1/2 (S217/S222), ERK1/2, pERK1/2 (T202/Y204), FRA1, pFRA1 (S265), RSK1/2/3, pRSK3 (T356/S360), FOSB, EGR1, AKT, pAKT (S473), S6, pS6 (S240/S244), Cyclin D1, pRB (S780), N-cadherin, E-cadherin, vimentin, BIM, MITF, and GAPDH were purchased from Cell Signaling Technology. Antibodies against ERF and pERF (T526) were purchased from Invitrogen. Signals were developed using SuperSignal West Femto, Dura, or Pico chemiluminescent substrates (Thermo Fisher Scientific), and visualized using the ChemiDoc Imaging System (Bio-Rad).

RT-PCR reactions were performed using TaqMan Gene-Expression master mix (Applied Biosystems), FAM-labeled probe for DUSP6, and Beta-2-macroglobulin (B2M) as a normalization control. Samples were run on a 7900HT Real-Time PCR machine (Applied Biosystems), and data were analyzed and normalized according to the manufacturer's instructions (2^−ΔCt method).

Of all the cutaneous melanoma models in the Novartis PDX encyclopedia, 13 models had BRAF^V600 mutations (18). Figure 1A shows the common genetic lesions identified in these models. CDKN2A loss-of-function occurred at a higher frequency in these PDX models relative to human tumors (2), consistent with results from a previous study (30). MITF copy-number alterations and mutations were found in five PDX models with potential oncogenic effects (1), but no trends associated with drug response were observed.

The Novartis mouse clinical trial (18) demonstrated that response of the BRAF^V600-mutant PDX models to continuous treatment with encorafenib and binimetinib ranged from complete regression (CR), to partial regression (PR) and stable disease (SD; Supplementary Fig. S1). Consistent with the phase III COLUMBUS trial, where the disease control rate was >90% (4), none of the PDX models displayed outright resistance (ref. 18; Supplementary Fig. S1). In addition to the PDXs that were enrolled in the mouse clinical trial (18), one additional model, X-20767, was also included in subsequent experiments (Fig. 1B).

To identify factors associated with differential drug response in BRAF^V600-mutant melanoma in vivo, we characterized immediate and longer term transcriptional changes in response to encorafenib and binimetinib treatment in our PDX models. RNA sequencing (RNA-seq) was performed on tumor fragments harvested prior to treatment, 8 hours post a single dose, or 8 hours post the final dose on the ninth day of continuous treatment (Fig. 1C). Three biological replicates were collected for each time point, with the exception of X-1906, which only had two RNA-seq samples for day 9 due to limitation in tumor size. In total, two CR and four PR/SD models were profiled. The day 9 samples for X-2613 were excluded from the analysis, because they were made up of >80% mouse RNA.

We applied limma-voom, a well-established method for RNA-seq differential expression analysis (24, 25), to identify differentially expressed genes (DEG) between treated and untreated conditions for each model (Supplementary Table S1). For all six PDX models, more transcriptome reprogramming occurred after 9 days of treatment in comparison with a single dose (Fig. 1D). There was significant overlap in differential gene expression between early and late time points, suggesting that the early transcriptome changes were retained after prolonged treatment. Furthermore, only a small subset of DEGs were shared across multiple models, whereas a significant number of DEGs were unique to each model (Supplementary Fig. S2), indicating heterogeneity among PDXs.

In an attempt to better understand patterns of transcriptional reprogramming by BRAF and MEK inhibitors in vivo, we performed unsupervised clustering of treatment-matched samples from all 6 PDX models using DEGs from the aforementioned analyses (Supplementary Table S1). Only the top 1,000 most variable DEGs across all samples were included in the clustering analysis, which may enrich for model- and time-dependent gene-expression differences in addition to posttreatment transcriptome alterations. Five major clusters were identified (Fig. 2). To elucidate the biological functions of each cluster, gene set enrichment was performed using the Broad Institutes' Molecular Signatures Database (MSigDB) Hallmark (31) and MetaCore transcription factor gene sets (GeneGo; Fig. 2; Supplementary Table S2). Genes in clusters 3 and 5 exhibited potentially interesting model and treatment-dependent patterns, although we were unable to find any functionally meaningful gene sets enriched in these two clusters.

The expression of cluster one genes was reduced with treatment across all PDX models (Fig. 2). Multiple gene sets were enriched in cluster one, including NF-κB signaling, cell cycle and proliferation, and the activation of several transcription factors downstream of MAPK signaling (32). In addition, cluster 1 contained known ERK target genes including DUSP4, DUSP6, SPRY4, and CCND1 (ref. 32; Supplementary Table S3). The expression pattern, in addition to gene set enrichment results, indicated that these genes were downstream of MAPK signaling.

In contrast, the expression of cluster two genes increased with BRAF and MEK inhibition. Cluster two was enriched for the MITF activation gene set, which contained known MITF target genes such as MLANA, TYRP1, and BCL2A1 (ref. 12; Fig. 2; Supplementary Table S3). In addition, epithelial markers such as CDH1, SNAI2, and ZEB2 were also upregulated. This is consistent with the enrichment of a differentiated melanocytic transcriptional state during the early phase of drug treatment. Although this cluster was associated with MITF activity, many of the activated genes were not found in the in vitro–derived gene signature describing the high MITF, proliferative melanoma phenotype sensitive to MAPK inhibition (refs. 15, 33; Supplementary Fig. S3A). The upregulation of MITF transcriptional programming was described as a mechanism of drug tolerance, which protects melanoma against MAPK inhibitor treatment (17, 34). Based on known biology of MITF and the expression pattern of cluster 2, these genes appeared to be linked to adaptive response to MAPK inhibition in melanoma.

Genes in cluster 4 were highly expressed in CR PDX models at baseline, and downregulated with treatment. Gene set enrichment identified genes involved in epithelial–mesenchymal transition, NF-κB signaling, and the activation of multiple transcription factors known to be downstream of ERK (32). Although AXL was identified in this gene set, there was very little overlap between cluster 4 and the in vitro–derived gene signature for the invasive melanoma phenotype associated with drug resistance (refs. 15, 33; Supplementary Fig. S3B). Consistent with gene set enrichment results, cluster 4 contained genes representing a more “mesenchymal-like” phenotype, in addition to ERK target genes such as FOSL1 (FRA1), EGR2, and DUSP5 (32). These results suggested that the genes in cluster 4 defined a MAPK-dependent, “mesenchymal-like” phenotype associated with sensitivity to BRAF and MEK inhibition.

Taken together, hierarchical clustering revealed distinct patterns of gene expression that were dependent on MAPK inhibitor treatment, drug sensitivity, and baseline model variability. Unsupervised clustering using the top 1,000 most variable DEGs demonstrated that there are three major sets of genes related to canonical MAPK signaling, drug tolerance, and drug sensitivity, respectively. Whereas an MITF-dependent, differentiated melanocyte transcription program was induced upon drug treatment, the expression of a subset of MAPK-dependent genes featuring “mesenchymal-like” markers was enriched in melanoma PDXs that achieve complete response upon BRAF and MEK inhibition.

Using RNA-seq data from untreated tumors, we asked whether baseline expression of the identified clusters from Fig. 2 was correlated with response to encorafenib and binimetinib across all 13 BRAF^V600-mutant melanoma PDXs (18). Genes associated with the MITF-high drug tolerance phenotype were more highly expressed in PR/SD models relative to CR models at baseline (Fig. 3A). The calculated signature scores were statistically significantly different between the two groups (Fig. 3B). On the other hand, the CR models expressed elevated levels of drug-sensitivity genes in comparison with the PR/SD models (Fig. 3C and D). There were no significant differences in the expression of canonical MAPK signaling genes between CR and PR/SD PDXs (Supplementary Fig. S4A; Fig. 3E).

To validate our findings, we assessed relevant markers and pathways at the protein level (Fig. 3F). Several posttranslational and transcriptional targets of ERK were highly expressed and/or phosphorylated in CR models. These include FRA1 (FOSL1), fosB, and EGR1 (32), consistent with the enrichment of relevant gene sets in the drug-sensitivity cluster (Fig. 2). Furthermore, the complete responders exhibited increased mesenchymal protein markers, whereas the epithelial markers were highly expressed at the protein level in the less sensitive PDX models (Fig. 3F). The AKT–mTOR pathway was also profiled given its known role in BRAF^V600-mutant melanoma drug response (35, 36), though a clear trend was not observed.

We showed that in BRAF^V600-mutant melanoma PDXs, the MITF-high, “epithelial-like” phenotype is not only associated with adaptive response, but also reduced sensitivity to MAPK inhibition. Inversely, the expression of a MAPK-high, “mesenchymal-like” transcriptional program is linked to sensitivity to BRAF and MEK inhibitors. This finding is distinct from the MITF-high (proliferative) and -low (invasive) phenotypes described in melanoma cell lines, which highlights important differences between in vivo and in vitro experimental systems.

Next, we evaluated the ability of our gene sets to predict clinical outcome using a published cohort of 17 patients (29). Using a combined signature score made up of both drug tolerance and sensitivity genes, there were no differences in best response or progression-free survival between patients predicted to be sensitive and vice versa (Fig. 4A and B). However, analysis using on-treatment patient samples revealed that the expression of drug tolerance genes was uniformly increased during the course of treatment (Fig. 4C). The expression of drug-sensitivity genes during the course of treatment underwent heterogeneous changes among patients (Fig. 4D). The induction of the MITF-high, “epithelial-like” phenotype in all patients suggested that selection toward a more drug-resistant phenotype occurs during the course of treatment, consistent with our findings from PDXs described herein. Although the baseline expression of these genes was not associated with patient outcome, it could be due to numerous factors such as intratumor heterogeneity, small sample size, heterogeneous treatment regimen, or tumor microenvironment.

To further characterize gene-expression changes associated with tumor regression in vivo, we constructed linear models to identify DEGs posttreatment only in the CR, but not PR/SD PDX models (ref. 25; Supplementary Fig. S4B; Fig. 5A). More DEGs specific to the CR models were identified after 9 days of treatment in comparison with after a single dose. Gene set enrichment was performed using MSigDB Hallmark gene sets (31). Genes preferentially downregulated in the CR models relative to the PR/SD models were enriched for epithelial mesenchymal transition (EMT), cell cycle and proliferation, and MAPK signaling (Fig. 5B). Western blot revealed that these pathways were also inhibited in models that showed PR/SD in response to BRAF and MEK inhibition, but the changes were most significant in the complete responder X-1906 (Fig. 5C). Consistent with earlier gene-expression analyses, BRAF and MEK inhibitor treatment resulted in a shift toward a more “epithelial-like” phenotype, whereas two less sensitive models already exhibited high level of E-cadherin protein expression prior to treatment. Furthermore, the CR model X-1906 exhibited the highest level of BIM upregulation upon treatment, consistent with apoptotic induction and the observed tumor regression (37). Models that showed SD also upregulated BIM upon BRAF and MEK inhibitor treatment, indicating that this is not sufficient to yield tumor regression. Taken together, encorafenib and binimetinib treatment in BRAF^V600-mutant melanoma PDXs resulted in inhibition of MAPK signaling, cell cycle and proliferation, and transition toward a more “epithelial-like” cell state. CR in response to treatment is associated with strong inhibition of these pathways and upregulation of the proapoptotic protein BIM.

The gene expression and signaling studies suggested that MAPK pathway suppression was more robust in the CR versus PR/SD models. To further extend this observation, MAPK pathway suppression was measured in untreated and on-treatment tumor fragments (Fig. 1C) using ERK phosphorylation and DUSP6 mRNA (Fig. 6A; Supplementary Fig. S5). As an ERK phosphatase tightly regulated by MAPK output, DUSP6 expression is highly sensitive to MAPK pathway inhibition (32). Unsurprisingly, the complete responder X-1906 displayed markedly durable pathway suppression. In comparison, the PR/SD models exhibited varied levels of MAPK inhibition. Three of the five PDX models displayed either sustained MAPK signaling or signaling recovery after 9 days of treatment, whereas the remaining two PDXs showed durable MAPK pathway suppression. This finding was consistent with on-treatment patient samples (29), where MAPK inhibition as measured by DUSP6 mRNA was greater in responding tumors but more variable in patients with less tumor shrinkage (Fig. 6B). In addition, AKT phosphorylation exhibited variable changes upon treatment, highlighting intertumor heterogeneity in adaptive response to MAPK inhibition (Supplementary Fig. S5A). These findings indicated that BRAF^V600-mutant melanoma elicited both MAPK-dependent and -independent adaptive mechanisms to survive under MAPK inhibitor therapy.

To evaluate the hypothesis that BRAF^V600-mutant melanoma varies in their dependence on the MAPK pathway for survival, we assessed the efficacy of triple combination with encorafenib, binimetinib, and the ERK inhibitor Vx-11e (38). The three-drug combination was well tolerated in vivo (Supplementary Fig. S6). We evaluated the triple combination in two PDX models that exhibited either MAPK signaling rebound (X-20767) or continual pathway suppression (X-4849; Fig. 6A). In the MAPK-activated model, combination with the ERK inhibitor resulted in further tumor regression in comparison with BRAF and MEK inhibitors alone (Fig. 6C and D). The ERK inhibitor Vx-11e stabilizes the phosphorylated form of ERK; therefore, MAPK pathway inhibition was assessed by the phosphorylation of ERK substrate RSK3 (32). Tumor samples collected after 9 days of treatment showed that the triple combination resulted in increased suppression of MAPK signaling, in addition to the upregulation of BIM (Fig. 6D). In contrast, triple combination with an ERK inhibitor did not improve antitumor activity in X-4849, which did not demonstrate MAPK reactivation following BRAF and MEK inhibitor treatment (Fig. 6E). This was despite the triple combination yielding deeper suppression of the MAPK pathway, as evidenced by reduced RSK3 phosphorylation relative to BRAF and MEK inhibitor double combination (Fig. 6F). Consistent with the lack of tumor regression, we did not observe any BIM upregulation upon treatment. To identify additional combinations that may be efficacious for MAPK-inactivated PDX models, we assessed efficacy results from the Novartis mouse clinical trial (18). For model X-4849, combination of encorafenib and the PI3K inhibitor burparlisib resulted in markedly more tumor regression in contrast to treatment using encorafenib and binimetinib (Fig. 6G). This suggests that targeting parallel pathways rather than vertical combination could improve the efficacy for a subset of BRAF^V600-mutant melanomas that do not reactivate the MAPK pathway during treatment.

As shown, these data support the notion that some BRAF^V600-mutant melanomas do not depend on reactivation of the MAPK pathway to survive under BRAF and MEK inhibitor therapy. Inhibition of the MAPK pathway could only achieve limited efficacy in these tumors, and additional work is required to identify treatment strategies with curative benefit. On the other hand, triple combination with BRAF, MEK, and ERK inhibitors presents as a viable treatment strategy to improve response in BRAF^V600-mutant melanoma that depend on the MAPK pathway for survival, where reactivation of the pathway occurs as early as 9 days following treatment in vivo.

Here, we demonstrated that distinct transcriptional programs are associated with differential sensitivity to MAPK inhibition in BRAF^V600-mutant melanoma PDXs. Melanoma PDXs with MITF-high, “epithelial-like” gene expression were less sensitive to BRAF and MEK inhibition. PDX models with increased levels of MAPK signaling and expressing a “mesenchymal-like” transcriptional program were preferentially sensitive to MAPK inhibition. In addition, the lack of tumor regression in response to BRAF and MEK inhibition was associated with adaptive responses consisting of MAPK reactivation or alternative survival mechanisms, which in turn dictated their response to further suppression of the MAPK pathway. This study highlighted the intertumor heterogeneity in BRAF^V600-mutant melanoma, and the relevant challenges in optimizing therapeutic strategies for individual tumors.

In this study, we observed that elevated MITF activity and the “epithelial-like” transcriptional program were associated with therapeutic resistance rather than sensitivity in vivo. This finding contradicts in vitro studies, where the “MITF-low” invasive cell state is linked to intrinsic resistance to BRAF inhibitors and vice versa (12, 13). The reasons for such in vitro-in vivo discrepancies remain unclear. Although tumor samples from melanoma patients could be classified into proliferative and invasive subtypes (39), there were no obvious correlations between melanocyte lineage transcriptional programming with clinical response (40). Intratumor heterogeneity could affect biological interpretations in vivo. Melanoma cells switch between proliferative and invasive phenotypes in vivo (41), and these two cell types coexist in patient tumors (42). Moreover, the transcriptional and posttranslational regulation of MITF is highly complex and can be modulated by the in vivo tumor microenvironment (17). MITF plays pleiotropic roles under diverse cellular contexts (17), sometimes demonstrating opposing effects on melanoma resistance to targeted therapy (13, 43). In addition, downregulation of AXL expression was identified in BRAF inhibitor–resistant patient tumors (44). It is conceivable that both extremely high and low levels of MITF could protect against the cytotoxic effects exerted by MAPK pathway inhibition.

Low levels of MITF and increased “EMT”-like transcription signature in the CR PDX models could also be a direct consequence of MAPK pathway output. ERK and its substrate p90RSK1 posttranslationally phosphorylate MITF (45). Increased MITF expression and activity occurs shortly after BRAF inhibition, which induces a drug-tolerant state preceding disease progression (34). Although EMT has long been implicated in therapeutic resistance across a variety of cancer types and therapeutic modalities (46), melanocytes originate from the neural crest and therefore do not undergo true “EMT.” Oncogenic BRAF was shown to modulate NF-κB activity (47) and elicit EMT transcription factor reprogramming in melanoma (48). In an in vivo murine melanoma model, treatment with MEK inhibitor resulted in transition toward a more epithelial phenotype (49). Furthermore, the induction of EMT by the MAPK pathway and its effector FRA1 was described in several epithelial cancers (50–52). In support of this interpretation, we observed that elevated baseline MAPK signaling, in addition to high levels of FRA1 expression and phosphorylation, occurred in association with elevated transcriptional programs such as EMT and NF-κB. BRAF and MEK inhibitor treatment resulted in a shift toward more “epithelial” melanocytic program in PDXs, which could either be the consequence of MAPK pathway inhibition or selection toward more resistant populations. In patient samples, the increased expression of melanocytic antigens in response to MAPK inhibition is concomitant with elevated CD8+ T-cell infiltration (53). Immune evasion is frequently associated with acquired resistance to BRAF inhibitor–based therapy, as immune modulation plays a significant part in BRAF inhibitor–mediated efficacy in vivo (44). The potential synergy and robust single-agent activity make the combination of targeted therapy with immune-checkpoint inhibitors an incredibly promising therapeutic strategy for BRAF^V600-mutant melanoma (54). To this end, a major limitation of PDXs is the inability to model the tumor immune microenvironment. Current efforts in developing humanized and syngeneic mouse models will help address this gap in recapitulating human tumor biology.

Interestingly, analysis using a curated ten-gene MAPK signature demonstrated that high MAPK pathway activity was correlated with improved response to BRAF inhibition in BRAF^V600-mutant melanoma patients (55). This is consistent with our finding that PDXs with elevated MAPK pathway activity responded better to BRAF and MEK inhibitors. However, we were unable to demonstrate the predictive value of our gene sets using a publicly available clinical data set. This could be due to intrinsic variability in small clinical data sets, or additional factors such as the immune system or disease severity that were not captured by the PDX system. Larger preclinical and clinical data sets will be required to further elucidate the links between basal MAPK pathway output, MITF transcriptional activity, and clinical response to BRAF and MEK inhibitors.

Analysis of on-treatment PDX tumors revealed that response to BRAF and MEK inhibition involved significant downregulation of cell-cycle and proliferation genes. This is expected as the gene-expression changes corresponded to sustained inhibition of MAPK signaling. However, robust MAPK inhibition was also observed in a subset of PR/SD PDX models, indicating codependence on alternative signaling pathways and necessity to target these parallel pathways to achieve tumor regression. PTEN deficiency has long been associated with resistance against MAPK inhibitors in BRAF^V600-mutant melanoma (35, 36, 56), although PTEN loss-of-function alone was insufficient to predict MAPK dependency in the PDX models. On the other hand, three-drug combination of BRAF, MEK, and ERK inhibitors achieved tumor regression in PDX models that reactivated the MAPK pathway in response to BRAF and MEK inhibitors alone. It was shown that the triple vertical combination could overcome the emergence of MAPK-reactivated therapeutic resistance (57). Our data suggest that early adaptive response to MAPK inhibition may be an important predictor of response to triple combination of BRAF, MEK, and ERK inhibitors. To this end, profiling of early on-treatment biopsies may serve as a patient stratification strategy for different BRAF and MEK inhibitor–based therapies. Ongoing clinical trials of BRAF and MEK inhibitor–based triple combinations will help identify additional therapeutic options for this heterogeneous and aggressive cancer.

In summary, we presented a detailed pharmacodynamic characterization of BRAF^V600-mutant melanoma PDXs treated with BRAF and MEK inhibitors. The in vivo model system revealed patterns of therapeutic sensitivity distinct from in vitro studies, which underscores the necessity for diverse preclinical experimental systems. Results from this study demonstrated the intertumor heterogeneity in metastatic melanoma, and the need for improved precision medicine approach to ultimately cure this aggressive disease.

T. Feng is a postdoctoral scholar at Novartis Institutes for Biomedical Research. J. Golji is an investigator at Novartis Institutes of Biomedical Research. F.C. Geyer is an investigator at Novartis Institutes of Biomedical Research. H. Gao is a senior investigator at Novartis Institutes of Biomedical Research. J.A. Williams is an executive director at, reports receiving commercial research grant from, and has ownership interest (including patents) in Novartis Institutes of Biomedical Research. D.D. Stuart is an executive director at, and has ownership interest (including patents) in Novartis Institutes of Biomedical Research. M.J. Meyer is a senior investigator at Novartis Institutes of Biomedical Research, is a Head, Discovery Pharmacology and In Vivo Biology at BMS and has ownership interest (including patents) at Novartis Institutes of Biomedical Research and BMS. No potential conflicts of interest were disclosed by the other authors.

Conception and design: T. Feng, D.A. Ruddy, D.D. Stuart, M.J. Meyer

Development of methodology: T. Feng, D.A. Ruddy, D.D. Stuart

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): T. Feng, A. Li, X. Zhang, J. Gu, J.A. Williams, M.J. Meyer

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): T. Feng, J. Golji, D.A. Ruddy, F.C. Geyer, J. Gu, D.D. Stuart, M.J. Meyer

Writing, review, and/or revision of the manuscript: T. Feng, J. Golji, J.A. Williams, D.D. Stuart, M.J. Meyer

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): T. Feng, X. Zhang, D.A. Ruddy, D.P. Rakiec

Study supervision: H. Gao, D.D. Stuart

We thank the PDX team, especially J. Green and A. Loo, for establishing and maintaining the Novartis PDX collection; L. Fan and O. Iartchouk for performing RNA sequencing; J. Korn and A. Johnson for bioinformatic support; and G. Caponigro, M. Niederst, V. Cooke, and J. Engelman for helpful discussions.

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.