DNA-Guided Precision Medicine for Cancer: A Case of Irrational Exuberance?

The identification of recurrent mutations in the genomes of cancer cells has enabled the development of a series of targeted cancer drugs. Use of these drugs requires the selection of patients who are most likely to respond through the identification of their corresponding genomic alterations. Indeed, most targeted cancer drugs mandate the use of a companion diagnostic test based on these genomic alterations to identify eligible patients. These genotype–drug response relationships are most readily explained by “oncogene addiction,” a term coined by the late Bernhard Weinstein to describe the critical dependence of cancer cells on oncogenic “driver” mutations (1). Thus, BRAF^V600E-mutant melanomas critically depend on increased MAPK pathway signaling, and hence such tumors respond well to selective inhibition of BRAF and/or MEK kinases (2–4). These and other encouraging results with targeted agents have sparked a number of histology-agnostic studies in which the presence of the oncogenic driver mutations is the sole selection criterion for treatment with the corresponding targeted agent in so-called basket studies (5–7). A first disappointment of these studies is that the number of “actionable” mutations (mutations for which a targeted drug is available as an approved indication or through a clinical trial) is relatively low. In a recent study at the MD Anderson Cancer Center, of 2,000 patients whose cancers were subjected to DNA sequencing, 789 (39%) were found to have an actionable mutation. Even though MD Anderson is one of the largest oncology centers in terms of clinical studies, only 83 of these 798 patients (some 4% of the total) could be treated in a genotype-matched trial targeting these alterations. Similarly, in the SHIVA trial, a genotype-matched “off-label” drug could be identified for only 40% of the patients. A second disappointment was that the responses of the patients treated with molecularly targeted agents, defined by progression-free survival, were no better than those of patients treated by “physician's choice,” leading the authors to conclude that the “off-label use of molecularly targeted agents should be discouraged” (7).

That the genotype alone may not predict drug responses reliably suggests that the tissue in which the cancer mutation occurs can be a major factor in determining response to therapy (8). As one example, BRAF^V600E-mutant colon cancers have modest responses to BRAF inhibition compared to melanomas harboring the same oncogenic driver mutation (9, 10). This discrepancy can be explained by the expression of the EGFR gene in colon cancer, a receptor that is activated upon BRAF inhibition in these cancers (11, 12). On the other hand, PARP inhibitors are active not only in BRCA-deficient breast and ovarian cancers but also in BRCA2-deficient prostate cancers, indicating that some genotype–drug response relationships have only limited tissue dependence (13).

Responses to actionable mutations may also be limited because the term “actionable” is used rather loosely in the context of treatment selection. For instance, KRAS mutations are frequent in human cancer and are generally considered to be “actionable,” as the oncogene addiction model predicts that treatment of such patients with a MEK inhibitor (a kinase downstream of KRAS) would deliver clinical benefit. However, in both preclinical models of KRAS-mutant cancers and in early clinical studies, responses of KRAS-mutant tumors to MEK inhibitors have proven to be modest (14–16). It will therefore be important to define more unambiguously what the term “actionable” means. These examples highlight the notion that the genotype alone may, in many cases, not be sufficient to choose a targeted therapy for a patient and point toward the need to consider additional factors beyond the actionable mutations to select patients for targeted therapies.

As stated above, the differing responses of BRAF-mutant colon cancers appear to be due to differences in EGFR expression, pointing to the relevance of the gene expression context in which a mutation is present. Even within a single cancer type, drug responses can be highly variable, depending on gene expression patterns. For example, using hierarchical clustering analysis, three major intrinsic subtypes of breast cancer can be identified: basal, luminal, and HER2 type. Although one would expect all HER2-amplified tumors to be of the HER2 type, a significant fraction of the HER2-amplified breast tumors are of the luminal subtype by gene expression analysis. Importantly, the neoadjuvant responses of these luminal-type tumors with HER2 amplification to chemotherapy and trastuzumab are much less impressive than the HER2-amplified tumors that are of the HER2 type by gene expression also. This highlights that even within cancers of a single tissue of origin, gene expression patterns can modulate responses to targeted drugs (17).

Along similar lines, it has been shown that loss of the BRCA1 or BRCA2 tumor suppressor genes confers sensitivity to drugs that inhibit PARP (18, 19). Based on these observations, olaparib is now an approved drug for BRCA-mutant ovarian cancers. Genome-wide copy-number analysis studies have shown that BRCA1-mutant breast tumors have a distinctive pattern of copy-number gains and losses. Interestingly, the group of breast cancers that has this distinctive pattern of genomic aberrations includes a number of tumors that lack a BRCA1 gene mutation and are referred to as BRCA1-like. Such BRCA1-like tumors are very sensitive to agents inducing DNA double-strand breaks, an established property of BRCA1-mutated tumors (20). It will therefore be of considerable interest to investigate whether such BRCA1-like tumors lacking a mutation in the BRCA1 gene itself will also be sensitive to PARP inhibitors.

The above-mentioned example indicates that cancer development can display convergent evolution, leading to a situation in which the resulting phenotype of tumors in terms of gene expression is similar, but the underlying genotype is different. This notion is also echoed in studies of colon cancer. Microsatellite instability (MSI) in colon cancer is associated with a higher mutation rate and a good prognosis. A gene expression signature trained on MSI colon tumors also identified a group of colon cancers that are not MSI by the clinical diagnostic assay, but resemble MSI tumors in terms of gene expression pattern. Importantly, such MSI-like tumors also have the higher mutation rate and the good prognosis of true MSI tumors (21). These observations are of interest in light of recent findings that MSI colon cancers have good responses to checkpoint immunotherapy (22). As a final example, BRAF-mutant colon cancers have a distinctive gene expression pattern and have a poor prognosis upon relapse. However, some 10% of colon cancers do not have a BRAF mutation but share the gene expression pattern of BRAF-mutant tumors and also have a poor prognosis after relapse (23).

In summary, these studies highlight two important points. First, tumors having the same oncogenic driver mutations can differ significantly in their responses to targeted cancer drugs. Second, tumors that lack a specific oncogenic driver mutation may nevertheless display very similar responses to therapy due to similarity in gene expression patterns. These two findings may explain why basket trials, in which patients are matched based on the genotype of the tumor only, are to date only moderately successful.

Although there is much excitement about targeted therapies, chemotherapy will remain the mainstay of drug-based cancer therapies for the foreseeable future. One of the main drawbacks of the use of chemotherapy is that it remains difficult to predict who might benefit from such therapies. Two large-scale drug-screening efforts using cancer cell line panels have failed to identify recurrent genomic alterations associated with responses to chemotherapeutic agents (24, 25). Gene expression analyses may also offer solutions here. In colon cancer, gene expression–based clustering analyses have identified multiple intrinsic subtypes, of which the mesenchymal subtype has the worst outcome (26). Unfortunately, these mesenchymal tumors also show poor responses to chemotherapy and EGFR-based therapies (27, 28). There are no mutations unique to mesenchymal colon cancer, but such tumors can be readily identified through transcriptome analyses. That gene expression–based subtyping can identify subgroups having different responses to chemotherapies is also seen in glioblastoma (29) and in breast cancer, where basal cancers are more chemotherapy-responsive than luminal tumors (17).

The first studies using genotype-directed patient selection in nonapproved indications suggest that we may have been overly optimistic about the contribution of cancer genotype analysis to the prediction of responses to therapy. It is likely that prospective studies that include transcriptome profiling in addition to mutation/copy-number variation analyses will yield important new insight into the interplay between the mutations in the cancer genome and the modulation of these effects by patterns of gene expression. Such studies combining DNA and RNA analyses in relation to patient outcome are urgently needed to improve response prediction in oncology, but are hampered by a lack of clinical standards in performing genome-scale transcriptome analyses in a clinical diagnostic setting. Large cancer centers increasingly use next-generation sequencing of cancer gene panels to genotype cancers in a routine diagnostic setting, and the FDA has taken an active role in the regulatory oversight of these DNA-sequencing technologies. However, the variable use of germline sequencing as an intrapatient control, the use of patient material without quality control, and the use of different bioinformatics pipelines make it difficult to compare studies. Standardization will be essential for future data comparison and sharing.

Even more challenging is comprehensive gene expression analysis of cancers through transcriptome sequencing (RNA-seq). It is hardly used outside a research setting for diagnostic purposes, and a highly standardized protocol for isolation of RNA and performing RNA-seq in a quality-controlled diagnostic setting is lacking. One additional limitation of RNA-based diagnostics appears to be that, unlike DNA, cell-free tumor RNA cannot be analyzed from the blood of tumor-bearing patients. However, a recent study suggests that blood platelets accumulate RNA molecules from cancer cells, enabling transcriptomic analyses for diagnostic purposes potentially using as little as a single drop of patient blood (30). Such technologies, when further validated, may be valuable to further elucidate the interplay between gene expression and oncogenic driver mutations in cancer.

Although DNA sequencing has kick-started an era of precision medicine by providing biomarkers of response to targeted agents, we begin to realize that cross-talk between signaling pathways and gene expression patterns in cancers can obscure simple genotype–drug response relationships. RNA sequencing represents a first additional dimension to understand better how tumors respond to treatment, and its implementation in clinical studies should be a priority of the translational research community.

R. Bernards is founder and employee of Agendia and has ownership interest (including patents) in the same. No potential conflicts of interest were disclosed by the other author.

The work of the authors is supported by a Stand Up To Cancer-Dutch Cancer Society Translational Research Grant (SU2C-AACR-0912). Stand Up To Cancer is a program of the Entertainment Industry Foundation administered by the American Association for Cancer Research.