Targeting melanoma by small molecules: challenges ahead

Summary

Significant progress has been made in targeting melanoma using small molecule inhibitors, but challenges remain. Here we describe the history of screening approaches in melanoma and their limitations. We propose several approaches to refine our screening models to enhance the discovery process. It is hoped that this discussion will stimulate further improvements in our development of small molecule inhibitors for treatment of melanoma patients.

Introduction

The landscape of melanoma therapeutics has evolved dramatically over the last five years. While unprecedented advances have been made in the treatment of metastatic melanoma, over 75,000 patients are diagnosed and nearly 12,000 patient die yearly of the disease in the United States—an increase more than any other cancer type (Siegel et al., 2012). Despite unprecedented investment by pharmaceutical companies and academia in discovery of potential new drug candidates, incorporating ever more sophisticated screening approaches, the number of highly efficacious drugs approved has remained disappointingly few. Here, we briefly review the history of drug discovery in melanoma and describe major challenges and approaches to overcome them. We focus mostly on the hurdles in drug discovery, because the difficulties of preclinical and clinical development have been described elsewhere (Pammolli et al., 2011; Paul et al., 2010) and are beyond the scope of this discussion.

History of drug screening in melanoma

Efforts to develop drugs that preferentially kill cancer cells are not new. The development of cytotoxic chemotherapy, the mainstay of cancer therapeutics until relatively recently, began in earnest in the 1940s, principally directed against hematologic malignancies. The mechanisms by which these agents work, such as damaging DNA or its synthesis or interfering with cell division can lead to broad toxicity. Although cytotoxic chemotherapy is highly efficacious in management of certain human malignancies, its limited efficacy as single agents treatments for most tumors led to development of new approaches for drug screening.

Until the development of the NCI-60 tumor cell drug screen in the 1980s, much of anti-cancer drug discovery focused on development of anti-leukemic drugs in transplantable murine models (Shoemaker, 2006). Transcriptional and functional profiling (for example, using shRNA screens) led to a greater emphasis on molecular target-oriented screening. In melanoma, an early success in melanoma from this approach contributed to the development of inhibitors of MEK for melanomas with BRAF(V600E) mutations (Solit et al., 2006). These approaches ushered the era of target-based screens, which incorporate genomic information. Although there has been a rekindling of interest in phenotypic screens, target-based approaches greatly outnumber phenotypic screens.

The growing list of validated targets led Paul Workman to describe the current state of drug development as a ‘second golden era’ (Workman, 2005). Target-based screens require molecular targets that are reliably (or at least identifiably) required for an essential aspect of the cancer phenotype, such as growth, invasion, or metastasis (Hanahan and Weinberg, 2011). Four target classes have been described: (i) genetic targets, such as somatic mutations in BRAF, (ii) synthetic lethal (see below) or oncogene non-addiction dependencies such as DNA repair in BRCA1-mutated breast cancer or metabolic dependencies, (iii) lineage-specific dependencies such as estrogen receptor in breast cancer or MITF in melanoma, and (iv) targeting of host-cancer cell interactions (Collins and Workman, 2006). Hypothesis-driven approaches based on molecular drivers of cancer have led to some successes such as ipilumimab. We emphasize the utility of unbiased, systematic screens here because it is likely that the enormous task of developing new targets (and especially, their combinations) may more efficiently be tackled by this approach rather than the ad hoc approaches of the past. We describe the major challenges in developing therapeutically useful screens for small molecules in melanoma, and we propose several strategies to overcome them (Table 1).

Table 1.

Challenges to melanoma drug discovery

Challenge	Possible solutions
Erroneous molecular targets	Emphasize use of primary human specimens, or short-term cultures Expand screening to large number of cell lines to reflect genomic heterogeneity of melanoma variants
Failure to recapitulate in vivo sensitivity	Incorporate screening using stromal tissues/factors Use of 3-dimensional equivalents In vivo screening
Failure to measure cytotoxicity rather than proliferation	Direct measurement of cell cytotoxicity
‘Non-druggable’ target	Inhibitors of protein-protein interactions Synthetic lethal screening
Failure to consider or measure therapeutic index	Computation models Lineage-restricted targets Selectively targeting oncoproteins over wild-type protein
Acquired resistance	Combinatorial screening Screening of resistant clones

Challenge 1: better screening models

A legacy of the NCI-60 effort was the use of in vitro cell models that measure the anti-proliferative effects of candidate compounds. Unfortunately, such in vitro models may not adequately model melanoma for several reasons. First, these cells may not reflect the primary genetic profile of melanoma patients, or fail to account for the genetic diversity of melanoma (Sharma et al., 2010a). For example, β-catenin mutations are relatively common in cultured melanoma cell lines (Rubinfeld et al., 1997), whereas they are extremely rare in primary specimens (Rimm et al., 1999). Additionally, BRAF mutant melanomas appear easier to grow in vitro than NRAS mutant or BRAF/NRAS-wild-type melanomas, leading to a representation bias in the cell lines that are screened (Colombino et al., 2012; Davies et al., 2002; Lin et al., 2008). Although systematic identification of genetic alterations in melanoma such as The Cancer Genome Atlas Project is being performed on primary patient samples, most drug screens are still conducted on cell lines that have been propagated in long-term culture. Similarly, most cell-based screens have thus far relied on cell lines from easily biopsied metastatic sites, which may differ genomically from primary samples and from other metastatic sites (Navin et al., 2011; Yachida et al., 2010). Single nucleotide polymorphisms may alter sensitivity of drugs through alterations in signaling or drug properties (Kim et al., 2010). Collectively, these issues suggest that much larger numbers of cell lines may need to be screened to reflect the diversity of genetic backgrounds. Two such studies incorporating hundreds of cell genomically annotated lines have been recently presented (Barretina et al., 2012; Garnett et al., 2012) which confirmed the genotype-selective efficacy of BRAF inhibitors and identified several new genetic-chemical interactions.

Most drug screens are conducted without stroma-derived soluble factors, which have recently been observed to promote resistance to targeted therapy such as BRAF inhibitors (Straussman et al., 2012). Stromal tissue may also affect melanoma cells through non-soluble factors. For example, these tissues provide a matrix that bi-directionally communicates with tumor cells (Mueller and Fusenig, 2004). The roles of other non-autonomous features of melanoma, such as angiogenesis also cannot be easily modeled by in vitro approaches. In addition, these models do not usually incorporate tumor heterogeneity. Small subpopulations of JARID1A-expressing cancer cells exhibit reversible drug tolerance (Sharma et al., 2010b), suggesting that variations in drug resistance may be regulated by the microenvironement and that such phenotype-switching may complicate therapeutic strategies. While the role of stem cells in melanoma is controversial, it may be necessary to screen drugs that target stem cells within tumors that might be required for complete disease eradication (Boiko et al., 2010; Gupta et al., 2009; Schatton et al., 2008). Three-dimensional culture approaches (Kenny et al., 2007; Kunz-Schughart et al., 2004), microfluidic devices designed to mimic microenvironmental gradients (Walsh et al., 2009) and screens using co-culture systems may be an approach to address these issues.

An additional concern with current screening models is the emphasis on proliferation endpoints rather than cytotoxic endpoints. BRAF inhibitors have relatively modest effects on cell death but strongly suppress cell proliferation, which may partly explain their limited ability to completely eradicate disease in patients (Koomen and Smalley, 2011). Similar caveats apply to other, less well studied cancer phenotypes that are indispensable to cancer pathogenesis (Hanahan and Weinberg, 2011).

Traditional allografting of human cancer cells into immunocompromised mice fails to recapitulate many aspects of tumorigenesis and do not adequately predict why some drugs fail or even why some drugs show remarkable success. Newer genetically modified mouse models with humanized immune systems may have greater utility. Recently, a zebrafish model was used to demonstrate that targeting dihydroorotate dehydrogenase (DHODH) cooperated with BRAF inhibition in melanoma (White et al., 2011). While these models may overcome some issues with current in vitro models, their applicability to human biology requires validation. Murine models, such as the Tyr::CreER; BRAF(V600E); PTEN^lox/lox or corresponding Ras models provide an approach to screen compounds for activity in genetically defined animals. However, how large screens of anti-cancer compounds or their combination could be efficiently conducted in vivo will require development of new approaches, such as pooled drug screens. Nevertheless, mouse models have been exploited to identify synergistic drug combinations that can mimic RAS inhibition (Kwong et al., 2012).

Xenograft experiments also fail to account for the increasingly recognized importance of immune surveillance mechanisms. Such studies may now be possible when mouse models such as the BRAF/PTEN model above are backcrossed to homozygosity in C57Bl6 mice.

Challenge 2: expanding our targetome

The successes of vemurafenib (Chapman et al., 2011; Flaherty et al., 2010), imatinib (Hodi et al., 2008) and other kinase inhibitors that target the c-KIT/BRAF proteins in melanoma and other cancer types have demonstrated the clinical efficacy of targeting oncokinases in melanoma. Enzymes such as kinases and proteases can be more easily drugged because they have catalytic pockets that can be chemically targeted with relative selectivity and feasible on/off-rate kinetics and be contained within molecules small enough to be systemically distributed and capable of intracellular delivery. However, the relative paucity of new kinase targets is illustrated by several recent, melanoma exome sequencing efforts (Berger et al., 2012; Hodis et al., 2012; Krauthammer et al., 2012) which failed to identify any new frequently mutated kinase targets.

Direct inhibition of protein-protein interactions by small molecules has been generally disappointing, presumably because the chemical interface required to disrupt such interactions would typically be larger. Some progress has been made in designing hydrocarbon-stapled helical peptides that appear to overcome issues of cell permeability, stability, and significantly increase affinity to the target. Such an approach has been used to directly inhibit the NOTCH transcription factor complex (Moellering et al., 2009) and the oncogenic transcription factor β-catenin (Grossmann et al., 2012). Similarly, targeting of bromodomains that are present in some transcription factors such as c-MYC may offer some ability to target transcription factors (Delmore et al., 2011). Similar approaches have been used to drug oligomeric proteins of the apoptotic BCL-2 family (Verdine and Walensky, 2007). It remains to be seen how generally applicable such approaches will be for a large number of putative non-oncogene targets.

It will be of interest to evaluate how generally applicable these approaches will be to other targets. One candidate target of particular interest for melanoma therapy is the micropthalmia transcription factor, MITF. MITF is a lineage-restricted basic-helix-loop-helix transcription factor that is recurrently amplified and mutated in ~15–20% of human melanomas, although many more melanomas remain dependent on its expression (reviewed in (Haq and Fisher, 2011)). MITF drives oncogenesis by promoting cell survival, proliferation and oxidative stress resistance. Direct inhibition of MITF has proven difficult but targeting upstream regulatory pathways such as the tyrosine kinase c-kit, TYRO3, and the Wnt pathway could be alternative approaches to target its activity (Haq and Fisher, 2011). Histone decetylase inhibitors have been shown to reduce the expression of MITF, block MITF-dependent growth and have proven to be clinically useful and safe in other cancer types (Yokoyama et al., 2011). Because MITF is post-transcriptionally modified by ubiquitnation, SUMOlyation and phosphorylation, targeting of these pathways may be possible.

Finally, synthetic lethal screens to identify non-oncogene dependencies have been proposed as an approach to target non-kinase targets (Kaelin, 2005). Synthetic lethality arises when a mutation in one gene (such as NRAS) leads to a dependency on another gene. Such synthetic lethal genetic screens identify combination of mutations in two or more genes that leads to cell death, whereas a mutation in one of these genes does not. Such functional screens have identified additional kinase dependencies, for example, in Ras-mutated cancers (Bommi-Reddy et al., 2008; Scholl et al., 2009), illustrating the utility of this approach. However, there have been only a few examples of clinical success (Fong et al., 2009) with this approach at present despite promising pre-clinical data (Bernards, 2012).

Challenge 3: predicting therapeutic index and off-target effects

Many drugs fail, not due to failure to identify precise targets, but due to toxicity associated with predictable or unpredictable (off-target) effects (Paul et al., 2010). The concept of oncogene ‘addiction’ implies that a target is more essential for the survival of melanoma cell than normal cells and is the basis for much of targeted therapy (Weinstein, 2002). However, it remains a significant challenge to predict therapeutic index for a given target or small molecule. Bioinformatic approaches to predict such effects have recently been proposed and may hold promise. Prospectively identifying therapeutic selectivity through refining computational models of signaling networks holds some promise for the future. However, model systems to functionally test therapeutic selectivity are urgently required. Targets that are restricted to the melanocyte lineage, such as MITF, or other targets that are restricted to tumor cells compared to host tissues could aid in the development of targeted therapy with decreased risk of toxicity (Garraway et al., 2005). While vemurafenib was developed to selectively target oncogenic BRAF (thereby achieving some tumor selectivity) (Joseph et al., 2010), it is clear that this class of BRAF inhibitors also impacts RAF dimerization in other cells (Heidorn et al., 2010; Poulikakos et al., 2010), thereby inducing unanticipated MAPK pathway stimulation within BRAF wildtype cells. Still, it may be possible to selectively develop inhibitors that only target mutant versions of oncoproteins.

Challenge 4: overcoming resistance

Treatment resistance can be categorized as intrinsic (de novo) or acquired. With an understanding of specific mechanisms of resistance to BRAF inhibitors of, it has been possible to develop new generations of drugs, or combinations of drug, that may overcome these resistance mechanisms (Druker and Lydon, 2000; Johannessen et al., 2010; Nazarian et al., 2010; Villanueva et al., 2010). Thus far, the identification of resistance mechanisms has largely utilized in vitro selection of drug-resistant cells (Villanueva et al., 2010) which could then be utilized for validation in patient specimens. Such approaches have led to significant success, but current and future efforts will screen melanoma cells derived directly from patient biopsies and compare pre-treatment (drug-sensitive) to post-treatment (drug-resistant) tumor specimens. Because melanoma commonly metastasizes to subcutaneous tissues that may be amenable to biopsy, there is a strong opportunity for melanoma drug discovery to become a paradigm for this approach. Finally, the history of cancer chemotherapy strongly suggests that single agents alone will not suffice to eradicate solid cancer cells. Not only does targeting individual molecules have limited efficacy, monotherapy may also suffer from inferior toxicity profiles at maximally tolerated doses.

Redundancies in signaling, cross talk, and adaptive mechanisms as well as the genetic heterogeneity of tumors discussed above suggest that multiple agents may be needed (de Bono and Ashworth, 2010; Lehar et al., 2008, 2009). Combination therapy, which may be synergistic or additive, requires exponentially greater throughput in screening efforts (Keith et al., 2005). For example, for 2000 compounds there are nearly 2 million possible binary combinations. Network analysis suggests that overcoming network robustness may require targeting several molecular vulnerabilities simultaneously, which significantly challenges current generation screening methods (Keith et al., 2005). Pooled screens have recently been described to screen synergistic combinations for treatment of HIV, and could be adapted to screen cancer cell lines (Tan et al., 2012). Recently, a genotype-phenotype combination screen screens has been completed in melanoma (Held et al., 2013), which demonstrated the utility of these approaches, although definitive evaluation of the results in additional pre-clinical (and clinical) scenarios is needed.

Challenge 5: incorporating alternative models of screening

Essentially, all the models described previously focus on development of drugs to defined molecular targets, but the opposite approach has is possible. For example, the Connectivity Map approach utilizes gene expression profiles to map molecular profiles to existing drugs (Lamb et al., 2006). This suggests it may be possible to link molecular signatures of ‘undruggable targets’ to drugs based on their respective effects on gene expression. Additionally, computational models of drug action, such as network or evolutionary models (Al-Lazikani et al., 2012), may be particularly useful in the development of combination therapies. While reliable prediction of genetic-chemical interactions in silico is not possible at present, such models may become more precise in the future.

Conclusion

The development and success of therapies such as vemurafenib and ipilumimab have ‘raised the bar’ for new therapy in melanoma. Although we have clearly entered a new era of melanoma therapeutics, the frustratingly small fraction of patients who are cured by these drugs still requires our continued development of improved therapeutic approaches. Although we have made great strides already, the demand is growing, and challenges that we describe here will need to be overcome to maintain this momentum.