Nevins & West Join Forces
Researchers to Expand Gene Profiling in Breast Cancer
This article originally appeared in GenomeLife, Issue 2.
"Major advances almost always occur because of technology," says Joe Nevins, the James B. Duke Professor and Chair of Molecular Genetics and Microbiology and Director of the IGSP's Center for Applied Genomics & Technology. Still, having said that, he is the first to admit that his work on gene profiling in breast cancer has blossomed mainly because of his fruitful collaboration with Mike West, the Arts & Sciences Professor of Statistics and Decision Sciences. "It's really a completely cooperative, joint effort," Nevins explains. "We bring different strengths to it. We work together."
However the collaboration works, IGSP Director Hunt Willard is effusive about the results. "The unique combination of talents of these two guys and collaborators like [Professor of Medicine] Andrew Huang have already produced a succession of breakthroughs," Willard enthuses. "They are poster-boys for the type of cross-campus, interdisciplinary effort that is so common at Duke and so integral to the IGSP's vision."
The molecular side of the Nevins-West collaboration involves microarrays (aka "gene chips"), which are thumbnail-sized squares of glass dotted with thousands of DNA fragments each corresponding to an expressed human gene. In a typical assay, mRNA is extracted from a patient's tumor and allowed to bind to DNA arrayed on a chip. In the event the mRNA finds its corresponding gene on the chip, the resultant DNA-RNA combination will fluoresce under a laser light. Microarrays permit researchers to catalogue which genes are turned on and off in any given tumor, thereby allowing one to measure the activity of a tumor's entire genome in one instant. For the purpose of elucidating which genes are active in cancer, microarrays have become indispensable.
Gene Profiling: Individualizing Breast Cancer Risk
But why is gene activity important in cancer? To answer that question, one must first understand how things have been done up until now. The current state-of-the-art relies on two clinical factors to predict recurrence in breast cancer: lymph-node status and estrogen-receptor (ER) status. Whether a tumor has spread to the lymph nodes is considered the most important measure for predicting recurrence and overall survival. If the tumor has not spread, a patient's risk of recurrence is relatively low. If it has, the recurrence risk is higher. Such information is not merely academic, as higher-risk women typically embark on more aggressive courses of treatment involving higher doses of cytotoxic chemotherapies. Similarly, whether a woman's tumor is ER-positive is regarded as an important prognostic indicator. Some one-third of all breast cancers do not express ER; this condition is associated with a poorer clinical outcome.
What has become clear in clinical oncology is that placing women into prognostic categories based solely on these two measures is frustratingly inexact. On the one hand, clinicians can divide patients into two groups that are different from one another in a general way. However, within those two groups there exists a tremendous amount of heterogeneity. Thus, making accurate predictions in any particular case remains difficult. Indeed, one-third of patients with no detectable lymph-node involvement will experience a recurrence within ten years. "The whole problem," says Nevins, "is if you have ten women walk in the door with breast cancer, by and large you're faced with ten different diseases." The challenge, he says, is how to get at that heterogeneity and thereby understand the unique characteristics of the individual tumor and patient.
One key tool is DNA microarrays. What people tend to forget, say Nevins and West, is that whether or not a woman is lymph-node positive is really a result of the genes that are being expressed in her tumor. So why not just go straight to the genes? Nevins, West, Huang, and their collaborators have done exactly that using microarrays. By examining the expression patterns of multiple "metagenes"-collections of dozens of genes that tend to co-vary across a collection of tumor samples-the researchers reasoned that they would be able to quantify risk for each individual patient based on an aggregate of thousands of genes rather than simply being stuck with the two-pronged blunt instrument of "high" and "low" risk resulting from assessment of lymph node and ER status.
That's all well and good. But how can investigators accommodate and interpret such massive quantities of data? How does one sift through expression profiles of thousands of genes and correlate them with clinical outcomes? Nevins acknowledges that as recently as a few years ago, he was wholly unprepared for the microarray data deluge. "I was completely naïve about the role of statistical analysis in dealing with this kind of data."
Enter Mike West. For West, collaborating with Nevins was the product of both personal and professional chemistry. On the professional side, gene profiling was an area that struck West as intrinsically interesting, clinically relevant and one where he thought his skills might be useful. "As a statistician one has a license to work in anybody's backyard. That's why I became a statistician." And personally, "Joe and I just sort of hit it off. That was key."
West's approach was to use statistics to help develop what he and Nevins now refer to as "clinico-genomic models." These models represent statistical approaches that allow the team to incorporate not only huge amounts of gene expression data, but clinical and pathology data as well. This is one major respect in which the work of the Nevins-West group can be distinguished from others in the field. Thus, while virtually everyone working on gene profiling in breast cancer can stratify women into high- and low-risk groups, the Duke team seeks to individualize each woman's risk of recurrence. "We want to make the focus on the individual," West emphasizes. "Let's say for a given patient that our gene expression index stratifies her into a lower risk group. Well, let's look at her other indicators and ask the question: Can we evaluate probabilities of recurrence for individuals not only in the broader risk categories, but in much more refined risk categories?"
For the Duke team, the answer is a resounding yes. By adding in thousands of gene expression data points, the stratification process is repeated again and again, with each successive step refining a specific patient's risk profile. "Think of it as a tree," says Nevins. "You make an initial split: high risk versus low risk. Then you look at one of those two groups and find another expression pattern that splits that group into low risk and high risk again. Keep on doing that and in essence what you're doing is teasing apart that heterogeneity." Eventually, one arrives at a probability for a particular outcome that has been custom-made for each patient.
Taking the Next Step: Better Metagenes, More Patients
Thus far, this approach has been able to predict with about 90-percent accuracy whether a breast-cancer patient will experience a relapse within three or four years. However, the West and Nevins team want to improve upon even that high figure, eventually eliminating all of the equivocal cases. To accomplish that, they believe that they must address two issues. First, they must look for other ways to aggregate genes into metagenes. Nevins suggests that other properties besides simple covariance in expression patterns must be taken into account. One possibility: using known biological information as a basis for aggregation, such as grouping together genes that code for proteins known to reside in the same biochemical pathways. "It may be that we create a whole series of different ways of organizing the genes and then look for the overlaps among those different sets."
View larger imageThe other problem is the numbers game. To date, the Duke group has subjected less than 200 patients to clinico-genomic profiling. Ten to twenty percent of those fall into the equivocal group. Thus, while that's great news in absolute terms, it is difficult to study so few equivocals in order to draw conclusions as to why more reliable risk predictions cannot yet be made for them.
Both of these challenges have led the Nevins and West group to the next wave of plans for the breast cancer project. Among these is a planned expansion of gene profiling to a much larger set of patients. Such an expansion would provide opportunities to explore novel ways of constructing metagenes as well as larger numbers of equivocal patients to study in order to further evaluate their clinical, genetic and tumor pathology profiles.
Nevins takes pains to emphasize that slow and steady will win the race. "We're still in the research phase of building confidence and trying to better understand things. Taking existing samples, looking at retrospective information, building the model-that's what we've done thus far." He sees two ways to begin extending the work. One is to study a whole new set of samples completely independent of what was studied before and test how well the predictive models perform, that is, to perform clinico-genomic profiling research on a prospective basis. In other words, to bring the technique into the realm of genomic medicine.
The other, bolder step would be to organize a formal clinical trial. But how? To classify women as lower risk based on a novel technology and then withhold treatment from them on that basis would be ethically dubious at best and certainly not acceptable within the oncology community. Obviously, Nevins, Huang and West recognize this. Instead, they can envision at least one scenario under which a patient group identified as lower risk undergoes gene expression analysis. Those women who would be predicted to be at higher risk than what might have been gleaned without gene profiling would then be randomized into two groups: one group would receive the same treatment that they would have received otherwise, while the other group would receive a more aggressive treatment. Of course, the prospect of enrolling patients in such a trial under the assumption they would be randomized for potentially harsher treatment has potential ethical quandaries of its own, but it could nevertheless allow objective assessment of whether stratification based on clinico-genomic profiling can indeed accurately predict recurrence and prolong survival.
Ultimately, if the current data stand up, gene profiling will become a commercial enterprise. And yet, Nevins, Huang and West readily acknowledge that issues such as production and licensing-let alone sales and marketing-are all terra incognita. Not only would the commercial players be breaking new ground by offering so-called personalized medicine, but the US Food and Drug Administration is playing catch-up as well. "The FDA is learning by doing, too," says West. "A piece of statistical software and a computer do not constitute a medical test. They are not a 'medical device.' So, there are no [official] answers yet because nobody has those answers." In late September, the FDA admitted to the Genome News Network that while it is trying to learn more about microarray technologies through meetings, seminars, and scientific workshops, the agency is still very much on a learning curve. In one pilot project, Research Triangle Park-based Expression Analysis, Inc.-a privately-held firm upon whose Scientific Advisory Board both Nevins and West serve- is preparing a "mock submission" of microarray data for the FDA in order for the agency to better understand exactly how microarray data might be used in regulatory submissions.
New Strategies, Both Early and Late
Looking ahead, some of the most exciting developments on the Duke team's horizon are detailed in a grant application that was submitted in August. In part, the grant focuses more on early stages of breast cancer. Nevins explains, "Everything we've done thus far has concerned invasive breast cancer-malignant tumors that have not yet metastasized to distant sites but nevertheless have characteristics of invasive cancer. The question is: Can we begin to better understand the events associated with progression up to that point?"
Concentrating on early-stage disease will also allow the West-Nevins team to exploit gene expression analysis in a more conventional way: to simply identify genes involved in the early stages of disease. In addition, the investigators plan to scrutinize late-stage metastatic disease. By taking advantage of novel technologies for sampling previously inaccessible metastatic lesions in sites such as lung and liver, Nevins hopes he and his colleagues can bring personalized treatment to patients for whom little hope exists at present. The idea, he says, is to sample a patient's metastatic lesions, perform gene expression profiling on them, and then try to interpret the cellular events that have occurred in that patient's metastatic tumors in a way that suggests to her physician that one or another experimental drug might be especially appropriate.
Genomic Medicine: Finally Coming of Age?
West shares Nevins' excitement surrounding the team's current and future plans, particularly with regard to the promise of the early-stage work. "We know what can be done. The principles have been proven now, both by us and by others," he asserts. "Can we now imagine taking it back to the very early stages? If we can, then we're talking about not only more effective treatment for breast cancer but an approach to an eventual cure through early detection. That's not going to happen in the next five years. But we expect to be contributing to that process."
Nevins agrees that the best is yet to come; however, he does not want to minimize what the Duke team is doing in the near term. He points to the application of genomic profiling within the next year to better guide the decisions of treatment for those that already have breast cancer. "Seeing this work have such an immediate impact on patient care is truly satisfying and exciting."
Further Reading:
Huang E, Ishida S, Pittman J, Dressman H, Bild A, Kloos M, D'Amico M, Pestell RG, West M, Nevins JR. Gene expression phenotypic models that predict the activity of oncogenic pathways. Nat Genet . 2003 34:226-30.
Huang E, Cheng SH, Dressman H, Pittman J, Tsou MH, Horng CF, Bild A, Iversen ES, Liao M, Chen CM, West M, Nevins JR, Huang AT. Gene expression predictors of breast cancer outcomes. Lancet . 2003 361:1590-6.
Huang E, West M, Nevins JR. Gene expression profiling for prediction of clinical characteristics of breast cancer. Recent Prog Horm Res . 2003 58:55-73.
Nevins JR, Huang ES, Dressman H, Pittman J, Huang AT, West M. Towards integrated clinico-genomic models for personalized medicine: combining gene expression signatures and clinical factors in breast cancer outcomes prediction. Hum Mol Genet . 2003 12 Suppl 2:R153-7. Epub 2003 Aug 19.