There have been several important recent advances in our understanding of tumor genomic ecosystems, and these advances have interesting implications for drug discovery in oncology.
The Journal Nature recently published a large data set on gene mutations in 21 distinct tumor types (http://www.nature.com/nature/journal/v505/n7484/full/nature12912.html). Much of the data came from the The Cancer Genome Atlas (TCGA) database, with additional data generated by the study authors. This study is sufficiently powered to uncover significance in several different ways. There is a cluster of mutations that are significant only in the combined tumor analyses, that is, when lumping different tumor types together. Conversely there a large cluster of mutations that are significant only in the analysis of individual tumor types, that is, the significance is lost if you look too broadly. Therefore these are genes that are important for specific tumor types. Finally there is a large cluster of gene mutations that are significant in both the combined analyses and in individual tumor analyses. This complexity of analysis is nicely shown in Figure 3 (http://www.nature.com/nature/journal/v505/n7484/full/nature12912.html#f3).
I spent a fair amount of time staring at this figure and going through the supplemental data (posted online and see also http://www.tumorportal.org/) and there are some results that I found interesting. First, the study confirmed many known cancer-related genes. The study also identified a fair number of new cancer-related genes mutated across or within tumor types, although these were found at the lower levels of significance. This is because they are mutated at a low rate, or the sample size for a particular tumor type was small, or both. The authors are transparent about this, and call for larger studies to increase sample size. This does beg the question as to the rate of gene mutation below which the knowledge is no longer actionable (because there will be so few patients), regardless the data will be critical to understanding tumor pathway biologies. Another interesting question is the extent to which new patterns of gene-mutation will emerge across tumor types, allowing binning (across tumor types) to complement subsetting (within a tumor type). Finally, the data might allow a different type of query, which is to ask which combinations of mutations are found within specific tumor types.
I want mention a few of the more common mutations, because these data held some surprises for me (although some readers know all this already, I’m sure). First, the best known cancer-related gene mutations cluster at the very highest levels of significance both across the 21 tumor types and within specific tumors. This makes sense, as these genes include those that contribute obligate cancer mutations: TP53; PTEN, PIK3CA and PIK3R1; KRAS, BRAF and NRAS; APC; EGFR, etc. There were a few genes in this category that surprised me, not so much because they made the list but because these at first glance appear more common than I had thought. GATA3 is a good example. Mutations in this gene are most commonly see in breast cancer but there are enough mutations in other tumor types to drive significance in the pooled tumor analysis, even though no tumor type other than breast is significantly associated with GATA3 mutations. Examination of the FTL3 data reveal a very similar pattern: mutations are significantly associated with acute myeloid leukemia (AML), as is well known, but also present in other tumor types, notably endometrial tumors and lung adenocarcinomas. When the mutational data across tumor type is pooled, significance is achieved. What are we do with such data? I think the answer perhaps is to simply know that these mutations can occur, and to look for them when typical mutations are missing in a given patient’s tumor. Such cataloging is of course the goal of personalized medicine. The other use of such data is to raise awareness of rare drug resistance mutations that may arise when targeting the major tumor oncogenic pathway in a particular tumor type. Many examples of this phenomena have been described (more on this below).
A different pattern emerges when we look at some other genes that are commonly mutated across tumor types but whose significant in these analyses is lower, due to a lower mutational rate. IDH1 is a good example here, having significant association with AML and glioblastoma multiforma, as is well known, but also with multiple myeloma (MM) and perhaps chronic lymphocytic leukemia (CLL). IDH2 is also most commonly associated with AML, but is present in colorectal cancer at “near significance” (love that fuzzy language). Notably, no other tumor metabolism genes appear in the analysis.
There are same gaps too I think. Looking at those genes that are significantly mutated only in a specific tumor type or types, we find some interesting genes. TGFBR2 has been described as a mutational driver in colorectal cancer, along with SMAD4. In the present analysis SMAD4 and SMAD2 are found to be significantly mutated in colorectal cancer, but the TGFBR2 mutation rate only reaches significance in Head and Neck cancer, although a few mutations do appear in the colorectal cancer data set used. Either the original studies are incorrect, which does not make sense biologically (TGFBR2 protein signals through the SMAD pathway), or this is an example of sampling error. Again, bigger data sets may be needed. Other tumor-type restricted patterns of gene mutation are very well known, such as EZH2 and CARD11 mutations in diffuse large B cell lymphoma (DLBCL). The CARD11 observation is interesting, as these mutations are associated with activation of MYD88, a gene known to be mutated in DLBCL and CLL.
There are lots of examples like these, and the data are easy to see and analyze: this is fun data to play with so have at it (see http://www.tumorportal.org/).
There is much discussion in the paper on new genes identified, and we’ll have to see how much of it is actionable at the drug development level.
That brings us to a different data set. If you go to the tumor portal you can sort by tumor type. Choosing melanoma, a highly mutated cancer, brings forth a whole spectrum of genes. Here’s a screengrab right from the tumor portal site (http://cancergenome.broadinstitute.org/index.php?ttype=MEL)
In the table above, blue refers to known cancer-related genes, red indicates genes whose function is relevant to cancer biology, and black are novel genes. As many readers know, BRAF mutations are the canonical melanoma oncogenic driver, signaling through the MEK/ERK pathway to drive melanoma cell proliferation, migration and metastasis. Antagonists of the BRAF and MEK proteins have emerged as the best line of defense against melanoma, but its a complicated fight. BRAF inhibitors were developed several years ago, starting with vemurafenib (Roche). Although BRAF inhibition induced responses in many melanoma patients, BRAF resistance mutants and MEK1 escape mutations evolve quickly and patients relapse. Common BRAF resistance mutations include V600E and V600K mutations that confer protection against the first generation drugs. Second generation inhibitors that target the resistance mutations were developed, such as dabrafenib (GSK). In addition, the MEK inhibitor trametinib (GSK, Japan Tobacco) was approved last year for use in treating melanoma. Several weeks ago the combination of these two drugs was granted accelerated approval for the treatment of advanced (metastatic) or unresectable melanoma that is positive for either mutation (V600E or K). This is a great example of cancer genetics=driven drug development in action.
However, other mechanisms of resistance are independent of BRAF mutational status because of additional MEK resistance mutations. These additional mutational strategies were discussed in a series of papers published online on November 21, 2013, in Cancer Discovery. These studies used tumor samples from patients that had relapsed after either BRAF inhibitor of dual BRAF/MEK inhibitor therapy. Mutations were found in the MEK1, MEK2, ERK1, ERK2 pathway and the PI3K, AKT1, PTEN pathway (PTEN is a negative regulator of PI3K signaling to mTOR and AKT1). The papers were reviewed in the January 9th issue of SciBx.
What does this single example tell us about the mutational landscape and drug discovery. First let’s note that some of the resistance mechanisms for melanoma do not show up in the proposed melanoma mutational landscape chart above, that is, these did not appear in the tumor ecosystem until that ecosystem came under selective pressure via drug treatment. This has 2 implications: the first is that the TCGA type overview of tumor mutations is just one source of data and following patients longitudinally as they experience therapy is another source of data. The other implication is that the mutational landscape contains putative additional mechanisms of escape at least in some patients. So using our melanoma example, we see in the table evidence of other potential escape pathways (NRAS, several checkpoint genes, and KIT stand out to me). So how many drugs will any individual patient need to keep a rapidly evolving melanoma under control?
The good news is that drug developers have taken notice and ERK1/2. AKT and PI3K inhibitors of various specificity are under development. The bad news I guess is that this is just one example of how complicated cancer therapy is likely to become. One good question not addressed here is how the immune checkpoint drugs will overlay with targeted therapies, for melanoma and many other tumors. Thats a question for another day.