© 2002 Elsevier

Expression patterns for 7 and 20 genes that were selected as discriminators of relapse versus continuous complete remission (CCR) for two types of acute lymphoblastic leukemia, T-ALL and hyperdiploid >50 ALL. (from E-.J. Yeoh et al., Cancer Cell 1:133–43, 2002.)

Countless things can go wrong in the complicated cell division process. Checkpoints fail, genomic instability increases, and when anarchy reigns, cancers spread. In trying to assess what is doing the damage and predict the damage yet to be done, doctors have an admittedly blunt set of tools to profile renegade cells. Histopathology and general prognostic indicators such as health, age, and metastatic spread do little to account for the breadth of tumor variety even within fairly specialized tumor types.

Since researchers first started toying with microarrays, they have considered the molecular profiling of tumors to be an important potential use. The genes that switch...


Twenty years ago, NKI pathologists began a frozen tissue bank, says Laura J. van't Veer, NKI's head of molecular pathology. With mRNA intact, along with the patient's clinical histories, the collection would prove ideal for gene-expression profiling. A group including van't Veer, NKI's René Bernards, and Stephen Friend, a Rosetta Inpharmatics founder, looked at a group of samples from 98 women younger than 55 who had primary breast cancers but no lymph node metastasis.1

Standards then did not require follow-up chemotherapy after surgery, and cancer returned in 34 patients. Today, approximately 95% of the original 98 patients would have received chemo in the United States, and 85% would have been treated according to Europe's St. Gallen standards, meaning that 55% to 65% of the patients would have needlessly undergone debilitating treatments. The group measured the tumors' expression levels for 25,000 human genes, and through various clustering algorithms, they narrowed down a field of 70 genes whose expression levels appeared to correlate with later metastasis.

These findings could result in lower unnecessary costs for women's healthcare and insurance. "We have lowered the number of patients in the so-called high-risk category toward 60%," says van't Veer. Friend, now a senior vice president at Merck (which owns Rosetta), adds that the presumed ability to predict a tumor's future was startling. "Even though you could look under the microscope and they would all look the same ... some have built into them programs to become aggressive," says Friend. The group later published the results from a study that applied the 70-gene panel to tumors from 295 patients with various clinical prognoses.

All the patients were assigned good or poor prognosis signatures.2 At 10 years those with poor prognoses had a 50% chance of remaining free of metastasis, as opposed to 85% for those with a good prognosis



© 2003 Nature Publishing Group

(A) The prevailing model of metastasis suggests that acquisition of metastatic capacity occurs late and is a rare event. (B) Gene expression data suggests that the capacity to metastasize might be acquired early and is shared by the whole tumor-cell population. (C) Additional findings support a model in which primary tumors with a high metasta-tic cpacity contain subpopulations of cells which display tissue specific expression profiles, predicting the site of metastasis (green, bone; blue, lung; yellow, liver). (from L.J. van't Veer, B. Weigelt, Nat Med, 9:999–1000, 2003.)

Researchers are hardly surprised that tumors are diverse, but clinicians don't have a singular approach to assessing those differences or identifying new cancers. Cancers with many heterogeneous classes such as leukemia have become a test bed for class prediction by microarray. A group including Todd Golub and Eric Lander at the Whitehead Institute and the Massachusetts Institute of Technology Center for Genome Research applied these technologies in 1999 to molecularly type ALL and acute myeloid leukemia (AML), which were known to be different.7

The group's 2002 Hot Paper followed up this work by looking at a less clear distinction in pediatric ALLs.3 Infants with leukemia don't respond to chemotherapy as well as older children, says Golub. "That could be because there's something different about [infants] and how they respond to drugs, or it could be that the type of leukemia they get is fundamentally different."

First author Scott Armstrong says a difference was expected, "but I think it was the magnitude of the difference that really surprised us and surprised everyone else." Based on their findings, they suggested that MLL be considered a distinct disease. Additionally, topping the list of active MLL genes was a receptor tyrosine kinase called FLT3, which has known small-molecule inhibitors; later work solidified the target for MLL.8 "Based on that, we're going to be starting a clinical trial with that FLT3 inhibitor," adds Armstrong.

James Downing of St. Jude Children's Research Hospital and others took a broader approach to ALL classification, chasing down the molecular signatures of more than six important leukemia subtypes, including MLL.4 He says: "We were able to show that there were clearly distinct gene-expression profiles for the known prognostic and biologic subtypes of pediatric ALL ... and that they could be used as a primary single-platform approach to accurately diagnose subtypes of pediatric ALL." Although there were hints that beyond subtyping, the expression profiles were giving useful information about leukemia biology and patient outcomes, Downing laments that his group's sample size (approximately 360 leukemias) was still not large enough. Some profiles hinted that gene expression might predict patients who would relapse or develop secondary AML as a result of chemotherapy. Subsequent work will explore this in larger studies, Downing adds.



© 2002 Massachusetts Medical Society

Hierarchical clustering of diffuse large-B-cell lymphomas from 274 patients according to expression of 100 DLBCL subgroup distinction genes. On the right are genes characteristically expressed in germinal-center B-cell-like (GCB) DLBCL or activated B-cell-like (ABC) DLBCL. The dendrogram at the top shows the degree to which each DLBCL is related to the others with respect to gene expression. (from A. Rosenwald et al., N Engl J Med, 346:1937–47, 2002.)

Predicting response to treatment for other diseases has been more successful. NCI's Louis Staudt, chief of the lymphoid malignancies section, has been looking to identify signatures that would predict the very different outcomes generally seen in diffuse large-B-cell lymphomas (DLBCL), the most common form of non-Hodgkin lymphoma and one that is known to have widely divergent survival rates.

In 2000 this group showed that such predictions were possible, and split DLBCL into two signatures, one for germinal center B-cell lineages, which had a good response to treatment, and one for a later stage of B-cell differentiation, which fared poorly.9 But, says Staudt, "We knew that that distinction built only part of the clinical outcome differences ... and we were eager to find other molecular features of the tumors that would more fully explain why people survived and were cured or not by combination chemotherapy."

In their 2002 Hot Paper, they reapplied the analysis to a larger group of 240 patients and began finding overarching themes in the signatures.6 Besides having expression patterns indicative of early B-cell differentiation, good-outcome tumors had signatures suggesting the presence of major histocompatibility complex class II protein and other immune response activity. "Again, it seemed like the host response to the tumor was influencing survival. Then when you put these all together in a multivariate model of survival, we did considerably better at predicting outcome," says Staudt.

Using a smaller sample, Golub and others in Boston also published on DLBCL in 2002.5 They distinguished between curable or fatal using unsupervised learning algorithms, which essentially cluster the expression profiles independent of other information. The outcome-predicting genes from Golub and Staudt's groups (less than 20 for each) barely overlap, but they both appeared to predict survival independently of the International Prognostic Index, the standard way to stratify lymphoma patients based on various characteristics. Despite differences in data, Golub says the message is the same and that Staudt's group is further along in validating a prognostic signature. "That's, I think, getting close to where one might introduce it into the clinical setting."

Indeed, Staudt says that Phase III clinical trials on diffuse lymphoma will start soon. The trials will use molecular profiling to discover survival determinants for two treatments being evaluated, he says. This past October, an NKI spin-off company named Agendia began offering microarray-based tests to predict breast cancer aggressiveness, and the results will be factored into treatment decisions. But while applications are in the works, not all are certain how these tests will play out in clinical settings. "Up until now, it's all been very well controlled," says Armstrong. Differences in sample quality, handlers, and other variables may cause discrepancies in a field often charged with having reproducibility problems.

And questions remains about the best diagnostic platform. Chips may work in the clinic but will require drastic changes in sample preparation. Other procedures to measure gene expression, such as quantitative PCR, might work in samples that have been formalin-fixed (the standard technique). "I don't think it's clear what platform will be the most practical ... to move into the clinical arena," says Downing. Staudt reasons that with tests such as qPCR, even relatively small predictor sets (between 10 and 50 genes) could become cumbersome and expensive compared to microarrays.

Michael Bittner, at the Translational Genomics Research Institute (TGen) in Phoenix, argues that the mathematics involved need to be honed. Hierarchical clustering techniques, he says, are idiosyncratic, which accounts for some of the data discrepancies between groups using similar techniques. Other problems include small sample size and an unpredictable subject. "Cancer ends up causing a lot of dysregulation, and it's evolving on the fly," says Bittner. "So, each cancer may have its own personal history that makes it partially concordant with some larger separation, but it's still got a lot of individual features. We don't know yet which of those are important."

Others proclaim the technique practically ready. "It's not the FDA, it's not the availability of the initial findings, nor is it going to be the clinicians or the patients who are going to be the rate-limiting step," Friend says. "What's really the issue now is the investment that's required by diagnostic companies to take these tests and make them available."

Interested in reading more?

Magaizne Cover

Become a Member of

Receive full access to digital editions of The Scientist, as well as TS Digest, feature stories, more than 35 years of archives, and much more!
Already a member?