Reading between the lines; understanding drug response in the post genomic era

Abstract

Following the fanfare of initial, often dramatic, success with small molecule inhibitors in the treatment of defined genomic subgroups, it can be argued that the extension of targeted therapeutics to the majority of patients with solid cancers has stalled. Despite encouraging FDA approval rates, the attrition rates of these compounds remains high in early stage clinical studies, with single agent studies repeatedly showing poor efficacy In striking contrast, our understanding of the complexity of solid neoplasms has increased in huge increments, following the publication of large-scale genomic and transcriptomic datasets from large collaborations such as the International Cancer Genome Consortium (ICGC http://www.icgc.org/) and The Cancer Genome Atlas (TCGA http://cancergenome.nih.gov/). However, there remains a clear disconnect between these rich datasets describing the genomic complexity of cancer, including both intra- and inter-tumour heterogeneity, and what a treating oncologist can consider to be a clinically “actionable” mutation profile. Our understanding of these data is in its infancy and we still find difficulties ascribing characteristics to tumours that consistently predict therapeutic response for the majority of small molecule inhibitors. This article will seek to explore the recent studies of the patterns and impact of mutations in drug resistance, and demonstrate how we may use this data to reshape our thinking about biological pathways, critical dependencies and their therapeutic interruption.

Following the fanfare of initial, often dramatic, success with small molecule inhibitors in the treatment of defined genomic subgroups, it can be argued that the extension of targeted therapeutics to the majority of patients with solid cancers has stalled. This is despite a drug development program that has prospered with fruitful pipelines producing effective and well tolerated compounds (Rubin EH, 2012). However, despite encouraging FDA approval rates, the attrition rates of these compounds remains high in early stage clinical studies, with single agent studies repeatedly showing poor efficacy (Ivy SP, 2010). In striking contrast, our understanding of the complexity of solid neoplasms has increased in huge increments, following the publication of large-scale genomic and transcriptomic datasets from large collaborations such as the International Cancer Genome Consortium (ICGC http://www.icgc.org/) and The Cancer Genome Atlas (TCGA http://cancergenome.nih.gov/). In parallel, advances in functional genomics have enabled characterisation of the phenotypic effect of many genes commonly mutated in cancer (Wan PT, 2004; Shalem O 2014).

However, there remains a clear disconnect between these rich datasets describing the genomic complexity of cancer, including both intra- and inter-tumour heterogeneity, and what a treating oncologist can consider to be a clinically “actionable” mutation profile. Our understanding of these data is in its infancy and we still find difficulties ascribing characteristics to tumours that consistently predict therapeutic response for the majority of small molecule inhibitors. This article will seek to explore the recent studies of the patterns and impact of mutations in drug resistance, and demonstrate how we may use this data to reshape our thinking about biological pathways, critical dependencies and their therapeutic interruption.

1 Genomics has opened the door to unexpected therapeutic gains in managing metastatic disease

In trying to appreciate a context for the clinical problem of drug resistance it is important to begin with the notion that sensitivity in treating solid tumours is a relatively novel paradigm. When previously treating metastatic non-small cell lung cancer or metastatic malignant melanoma, durable responses were the exception to the rule in the setting of cytotoxic chemotherapy (Eigentler TK, 2003; Reck M 2013). The concept of therapeutically relevant genetic subgroups in these elusive diseases has been driven by the discovery and annotation of somatic mutations uncovered by hypothesis-free genome screens (Stratton MR, 2011). Oncologists treating patients with these same tumour type are now habituated in the use of mutations detected in BRAF, EGFR, or indeed re- arrangements in ALK, and more recently ROS or FGFR1, to alter clinical management. This science has changed ideas held by a generation and has been rapidly incorporated into clinical practice (McDermott U NEJM, 2011).

Following on from the success of these examples, the expansion and exploration of large-scale gene screens by collaborative groups such as TCGA and ICGC has been therapeutically enlightening in many tumour groups. Although not universally acclaimed (Gabor Miklos, 2005) this approach of free exploration, usually of the coding genome, has enabled us a glimpse into the biology of some unusual cancers and often to be surprised by what is uncovered. It could be argued that without this approach we would not be trialling BRAF inhibitors in subtypes of haematological malignancies (hairy cell leukaemia) and in Langerhans cell histiocytosis, both of which harbour activating mutations in BRAF (Haroche J et al., 2013). In addition to providing novel tractable targets in these rare tumours, the large scale “genomic landscape” approach has given us arguably the most complete overview of the heterogeneity and critical signalling dependencies of more common tumours such as breast cancer (Stephens PJ et al., 2012). These findings are now being exploited in gene stratified therapeutic subgroup studies, such as I-SPY-2 (http://www.ispy2.org/) and FOCUS4 (http://www.focus4trial.org/), that hope to demonstrate survival benefits in these genetically stratified groups of patients.

Conversely, it is in the use of this technology to explore outlier sensitivities that perhaps more therapeutically relevant biological insights have been gleaned. A perennial question for oncologists is the puzzling phenomenon of the outlier, the patient who responds unusually well to a given therapy. In using genomics to address this question we have learned of examples such as mTOR pathway activating mutations in TSC1 that predispose urothelial tumour patients to unusually durable responses to mTOR inhibition (Iyer G 2012).

Not only has the field explored both common and unusual tumours and found tractable mutations, but the dissection of functional biology surrounding these mutations has enabled progress in therapeutic strategies to exploit the newly identified subgroups. Indeed, in the space of a few years the management of BRAF mutant melanoma has changed so considerably that the focus of debate has now shifted from questioning the complete absence of effective therapeutics, to the most effective manner in which to deploy the multitude of effective MAPK-directed therapies in order to best delay the emergence of drug resistance (Sullivan RJ, 2013).

2 The emergence of drug resistance is to be universally anticipated with all targeted therapeutics

Despite documented complete remissions and the successful prolongation of both time to progression and overall survival in cancers where pharmacological intervention with targeted therapies is standard of care such as ALK re-arranged non small cell lung cancer (NSCLC) and BRAF mutant malignant melanoma, resistance to treatment is universal and presents in the form of acquired or intrinsic populations of cancer cells that no longer respond to the therapy (Shi et al., 2014; Wagle et al., 2011). Several recent studies have focussed on these cell populations in order to identify the landscape of genetic and transcriptomic mechanisms of resistance in order to therapeutically circumvent or prevent them emerging.

3 The challenge of the long tail of resistance effectors and clonal heterogeneity

Several pre- and post-therapy biopsy studies have utilised whole exome or whole genome based sequencing approaches to discover resistance effectors in tumour populations that have progressed through therapy. These biologically distinct populations within the tumour retain a proliferative or anti-apoptotic advantage under the selection pressure of the relevant therapeutic and expand to form the bulk of the progressing tumour, as seen with the emergence of KRAS mutant colorectal cancers in cetuximab treated patients (Misale S 2012). Mutations can serve to either maintain signalling through the preferred oncogenic pathway (despite therapeutic inhibition) or indeed bypass that pathway altogether, and switch dependency to an alternate route (Table 1). In metastatic melanoma, for example, resistance to BRAF inhibitors in the clinic has produced examples that satisfy both of these mechanisms, with mutations in the MAPK pathway that maintain signalling in the face of BRAF inhibition as well as mutations in other pathways that allow the cells to break with their BRAF oncogene addiction (Turajlic S 2014, Van Allen, 2013).

Dissecting Resistance to small molecule inhibitors in signalling pathway contexts.

Such mutations can directly abrogate the function of the target gene itself (Table 1). One of the first studies of this kind in EGFR mutant NSCLC identified “gatekeeper” mutations at residue T790 that change the conformational state of the target kinase pocket, significantly affecting the binding kinetics of the EGFR inhibitor gefitinib (Inukai M 2006). Similar mutations in the ALK kinase domain, or copy number changes in the ALK gene, are seen in acquired resistance to crizotinib in ALK re-arranged NSCLC (Choi, 2010). Furthermore, mutations or copy number gains that result in high level re-activation of the target pathway are also seen in BRAF mutant melanoma where COT (MAP3K8), BRAF amplification and NRAS point mutations all result in sustained MAPK activation in the presence of BRAF and/or MEK inhibition (Rizos, 2014; Trunzer K 2013).

This approach has demonstrated conclusively that resistance can be mediated by somatic mutations altering either the pharmacokinetics of the drug itself or modifying the biological consequences of target inhibition. It is presumed that such mutations pre-date the introduction of therapy and evolve in a Darwinian fashion in the presence of the therapy as an evolutionary bottleneck. Intratumoural genomic heterogeneity has been conclusively demonstrated in clear cell renal carcinoma and presumed to be the substrate for this Darwinian emergence of drug resistant subclones in other tumour types (Gerlinger et al., 2012). Recently, exome screens comparing pre- and post-treatment metastatic tissue in melanoma have confirmed clonal expansion and branched chain evolution of the tumour under BRAF inhibitor selection (Shi H 2014). However, more recent data has shown intriguing evidence that metastatic lesions in ALK re-arranged NSCLC progressing under crizotinib treatment may indeed be derived from more than one subclone, and that there may be exchanges of tumour cells from non-contiguous tumour populations – termed “self-seeding” (Awad et al., 2013).

Such screens have also demonstrated an inherent weakness as they rely heavily on a priori knowledge to identify mutated genes previously annotated in oncogenic pathways. Sample number can be prohibitive in such studies to discover with high confidence novel resistance-causing variants that present with low frequency. The recent informative study by Van Allen (2014) exploring the clinical landscape to RAF inhibition in melanoma has demonstrated recognisable and recurrent resistance effectors in the majority of the 45 patient cohort, and has highlighted that many tumours fall into the category of harbouring low frequency putative resistance-causing variants such as HOXD8 and RAC1 that fall outside of the MAPK pathway. This has been termed the “long tail” of resistance effectors, and may explain why 20–30% of these patients have no recognisable resistance-causing variants, although it is worth noting that somatic mutations are not responsible for alternative splicing of BRAF, or ligand mediated resistance, all of which have been described in this context (Poulikakos et al., 2010; Wilson TR et al., 2012). A similar percentage is quoted in the proportion of crizotinib resistant ALK mutant patients with an unknown resistance mechanism (Doebele RC et al., 2012; Perez CA et al., 2014).

That clinical resistance studies are by necessity underpowered to detect recurrence in infrequently mutated genes poses a specific problem. It has been recently postulated that to detect significant but infrequently mutated genes that recur in under 2% of a tumour, a cohort would require between 900 and 2500 tumour normal pairs depending on the background mutation rate (Garraway and Lander, 2013). This number, while feasible for exploring the genomic architecture of primary tumours, will clearly be challenging to achieve in the context of drug resistance. There will therefore always be a need to have some a priori hypothesis of candidate resistance effector genes to examine in such resistance cohorts, a topic we explore below (Functional approaches exploring modifiers of drug response).

An additional challenge is that posed by recent findings that different drug resistant metastases from within the same patient can harbour not only different mutational mechanisms within the same oncogenic pathway, but also mutations that putatively activate alternate signalling routes. A recent exploration of drug resistant biopsies from a cohort of melanoma patients following treatment with a BRAF inhibitor demonstrated multiple resistance mutations in different genes within separate metastases in the same patient (Shi et al., 2014). In a number of patients mutations activating either MAPK or PI3K signalling were observed in different resistant metastases in the same patient. The implications of these findings are that multiple lesions would require to be biopsied to capture the complete resistance landscape in any single patient, and combinatorial therapy strategies would need to be employed. One potential solution to this problem would be the ability to detect a signal of all of the combined mutations across all metastases by capturing and sequencing circulating cancer DNA (Murtaza et al., 2013). As yet, it remains to be seen whether this is feasible.

4 The challenge of the uncharted genome

Exome based studies are also hampered by the fact that our knowledge of the role the non-coding genome may play in mediating drug resistance (and which represents 98% of the genome) is still in its infancy. There are recent examples that the parts of the genome that are not explored routinely in drug resistance studies reported to date (largely for reasons of cost) may harbour mutations that are relevant to therapeutics. Two specific nucleotides within the TERT promoter have been shown to be recurrently mutated in over 70% of melanomas and 16% of other tumours producing a novel binding site for the ETS transcription factor (Huang FW et al., 2013). As the number of whole cancer genomes available for analysis world-wide increases with falling sequencing costs, it is possible that recurrent mutations in these non-coding regions may attain increased significance in drug resistance cohorts.

Similarly while pleiotropic effects on gene regulation by miRNA have long been thought relevant to drug response, a recent study has specifically shown miRNA effects on the MET oncogene is a factor in resistance to EGFR inhibition in NSCLC (Garofalo et al., 2011). Similarly, miRNA on chromosome 9 that are lost with the tumour suppressor CDKN2A can have a marked effect on response to cisplatin in mesothelioma (Ivanov SV et al., 2010). Our knowledge of the significance of somatic mutations in miRNA genes and target binding sites is still incomplete but already it is clear that germline SNP's in the 3′ UTR binding sites of let-7a can affect the repression or activity of known oncogenes such as KRAS (Chin LJ, 2008). One can therefore speculate that somatic mutations in non-coding regions may play an even more complex and multi-layered role in determining the response to a therapeutic than previously acknowledged.

5 Functional approaches exploring modifiers of drug response

Given the problems and cost associated with uncovering low frequency resistance genes seen in the clinical studies performed to date, novel molecular biological approaches have been utilised to provide an alternative to exploring gene/drug interactions outside the clinical setting.

Forward genetic screens have demonstrated resistance mechanisms that clinical sequencing has not had the ability to detect. Loss of function screens using shRNA have demonstrated EGFR mediated feedback signalling to be a resistance mechanism to BRAF inhibition in BRAF mutant melanoma and colorectal cancer (Prallahad A 2012, Sun C 2014). Kinome wide siRNA screens in both melanoma and lung cancer demonstrated that NF1 loss mediates resistance to BRAF inhibition and EGFR inhibition respectively, in both tissues by activating Ras signalling (Whittaker SR, 2013; de Bruin EC et al., 2014). More recently, highly efficient gene knockdown has been achieved using CRISPR/CAS 9 mediated silencing which has independently demonstrated the role of NF1 and NF2 in resistance to BRAF inhibition (Shalem O et al., 2014).

In parallel, advances in gene-activating open reading frame (ORF) technology has enabled kinome- and genome-wide screens to identify genes whose expression (rather than loss) confers a change in drug response phenotype. One such screen in BRAF mutant melanoma has defined the important role of and MAP3K8 (COT) in mediating resistance to BRAF and MEK inhibition (Johannessen CM et al., 2010). A follow-up study has defined the role of G-protein coupled receptors, CREBBP and PKA as part of a lineage specific transcriptional programme in the resistance phenotype for MAPK directed therapies (Johannessen CM et al., 2013).

Such gain- and loss of expression screens in cancer have given us valuable insights into resistance mechanisms but are unable to capture specifically those point mutations which account for a number of clinically observed drug resistance mechanisms described to date such as KRAS^G12D. To this end the alkylating agent Ethinyl-nitroso-urea (ENU) has been used in murine and cell line based phenotype screens by utilising its propensity to cause point mutations stochastically throughout the genome. Crizotinib resistance studies in NSCLC cell lines have demonstrated not only which gatekeeper mutations in the ALK gene enable resistance to crizotinib, but have given an insight into resistance to second generation ALK inhibitors (Zhang S 2011).

The genetic perturbations induced by the means discussed above are susceptible to the criticism that such screens are by their nature artificial and may not reflect phenomena seen in vivo, alongside problems with the reproducibility of certain technologies (Babij C 2011). In this context several large scale cell line based screens seeking to annotate functional characterisation in terms of response to small molecule inhibitors with genomic and transcriptomic biomarkers have been reported recently (Barretina et al., 2012; Garnett MJ et al., 2012) The premise here is that cell autonomous genomic features that modify drug response are so myriad and multi-layered that one way of gaining insight into such subtleties may be in utilising a great number of features and input cell lines that harbour the same genetic alterations as their parent tumour groups. This effort has been transformative and has identified novel biomarkers as well as known and clinically validated gene–drug interactions (Garnett MJ et al., 2012).

As the technology that underpins such genome-wide perturbation screens improves, one can imagine a time in the very near future when all targeted agents are routinely tested through such screens to define the putative genetic landscape of drug resistance. Thus, when resistance does occur in the clinic there would be candidate genes to sequence in biopsies obtained upon disease progression and in all likelihood alternate therapeutics identified based on these findings.

6 The future

It is beyond doubt that we have entered the post–genomic age in cancer therapeutics. It is increasingly difficult to consider the care of patients with advanced disease without considering at the same time the molecular taxonomies which have become apparent through next generation sequencing. Classification remains crucial as we continue to move from cumbersome “all comers” late phase clinical studies to rapid moving smaller, enrichment type studies with the knowledge that critical dependencies are indicated by key mutational events and that these likely confer sensitivity to pathway inhibition. Furthermore the introduction of “ basket study” therapeutic trials, such as the recently announced MATCH Trial, where drugs are trialled in a gene rather than tissue specific context is likely to extend the scope of therapeutic intervention outside well established tissue groups (Willyard C 2013).

Beyond this initial stratification however we are learning to utilise these same ideas of a genetic taxonomy to construct a framework for dissecting resistance to treatment. This understanding has already borne fruit in enabling the introduction of second generation drugs effective against the most likely resistance mechanisms likely to occur in a particular context – as seen with Ceritinib in ALK re-arranged Crizotinib resistant patients (Shaw AT, 2014). Similarly, with the resistance landscape in BRAF mutant melanoma to BRAF inhibitors demonstrating a predilection of MAPK pathway re-activation, combination MAPK blockade was rapidly trialled improving time to progression and overall survival in BRAF mutant melanoma compared to single agent therapy (Flaherty et al., 2012).

We have arrived at arguably the most exciting time in cancer therapeutics for the last 30 years. Despite those technological advances that now enable one to sequence tumours with greater depth, speed and lower cost than ever before, we are at an inflection point where the philosophy of our approach is changing to reflect the biological questions we deem important. Rather than using genomics in isolation, we suggest that to continue to uncover and respond to resistance mechanisms, we will need to integrate functional approaches to enable the rapid characterisation of phenotype alongside genotype and utilise some of the many approaches discussed above to understand the consequences of somatic mutations in cell lines, animal models and ultimately patients. This will more than ever depend upon our ability to collaborate, and our willingness as a community to share data, resources and expertise.