|
||||||||
|
DISEASE PROFILING Diagnosis by Database Shows PromiseStudies Classify Cancers by Gene Expression, Reveal New Forms of Disease and Novel Approaches to Diagnosis They might as well have hung "Gone fishing" signs on their lab doors, some folks thought. A couple of years ago, cancer researchers in Boston went looking for distinctive patterns in thousands of genes turned on or off in diseased cells. They thought genetic profiles might help them diagnose, treat, and understand their patients' diseases better. The effort reminded some naysayers of Henry Fonda in the movie On Golden Pond, trying in vain to catch a fish named Walter, the big one who gets away.
Five teams of researchers have discovered new types of cancer or new ways to diagnose known cancers by analyzing gene expression patterns of thousands of genes in diseased cells. Pictured from left are Scott Armstrong, Sridhar Ramaswamy, Matthew Meyerson, and Todd Golub. Photos by Steve Gilbert Now, five teams of researchers report they have hooked--if not landed--some legendary lunkers. Genomics attracts as much hyperbole as a good fishing story. Even taking that into account, these papers suggest that genomics may deliver on many of its bold promises to medicine, at least in the arena of cancer. Fingerprinting TumorsIt turns out that simultaneous measurement of thousands of genes expressed in a tumor sample may, indeed, explain why some people respond to cancer treatments and others do not. Such transcriptional profiling may provide a more reliable diagnosis. It may uncover previously unrecognized kinds of cancer. It may provide new discovery tools to explore and ultimately to interfere with the basic machinery of cancer.For their study in the Nov. 13 online edition of Proceedings of the National Academy of Sciences, researchers looked for clusters of activity and inactivity in 12,600 genes from 139 lung adenocarcinomas and additional lung tissue samples. They found four previously unknown subtypes of lung adenocarcinoma, one of which carries a poor prognosis with a survival rate of 21 months--half the usual rate for the other three. The study was led by postdoc Arindam Bhattacharjee in the Dana-Farber lab of pathologist Matthew Meyerson, HMS assistant professor of pathology. Next, in the January 2002 Nature Genetics (posted online Dec. 3) another team using the same technique was surprised to find enough gene activity differences in a rare and aggressive form of acute lymphoblastic leukemia (ALL) to call it a unique form of cancer. They proposed the name mixed lineage leukemia (MLL). It strikes and usually kills infants in their first year, and it also appears to strike blood cells at an earlier stage of maturity than the two common leukemias. The project was led by pediatric oncologist Scott Armstrong in the Dana-Farber laboratory of Stanley Korsmeyer. Lethal PatternsTwo other research groups started with a different question: what is different about the cancers we cannot seem to cure? Genomics appears to predict patient outcome for diffuse large B-cell lymphoma better than the International Prognostic Index. One study was led by the researcher who developed the original index, Dana- Farber oncologist Margaret Shipp, an HMS associate professor of medicine. Some adults can be cured of this most common lymphoma, but most succumb to the disease. In the study technique, called "supervised learning," the computer is instructed to find the gene expression differences between patients who live and those who die.Using a similar strategy, another group found gene expression better than current clinical criteria for distinguishing childhood brain tumors and predicting responses of patients to chemotherapy and radiation treatment, especially medulloblastomas. The study was led by Children's Hospital pediatric neurologist Scott Pomeroy, an HMS associate professor of neurology. Preliminary results have been presented at meetings this year. Taking a more global view of diagnosing cancers by database, another group created a large expression profiling database spanning 14 of the most common human tumor types. Then they applied a new computational approach to molecular cancer diagnosis. For most samples, they got it right most of the time, according to their report in the Dec. 11 online edition of PNAS. That's not good enough for patients yet--pathologists usually get it right an estimated 98 percent of the time--but it showed molecularly complex cancers can be distinguished by gene expression profiling across classes, which may make this an additional tool in cancer diagnosis. This approach can be used to classify unknown tumors into newly discovered cancer types with differing prognoses. The study was led by Dana- Farber oncologist Sridhar Ramaswamy, HMS instructor in medicine. Toward Computing DiseaseBehind the scenes on all these papers are senior authors Todd Golub, HMS assistant professor of pediatrics; Eric Lander, director of the genome center at the Whitehead Institute of Biomedical Research and Massachusetts Institute of Technology; and a multidisciplinary, multi-institutional team that also includes mathematicians and computer scientists.Golub, who holds joint appointments at Dana-Farber and the Whitehead Institute, published the first "proof-of-concept" paper two years ago in Science. He and his collaborators used the basic tools--microarrays of gene activity and computer clustering algorithms--to distinguish two different types of acute leukemia--acute myeloid leukemia (AML) and acute lymphoblastic leukemia. "One could argue that proof of concept was like shooting fish in a barrel," Golub recently said. "Over the past 20 or 30 years, methods to distinguish AML from ALL evolved gradually by conventional means with high accuracy." Likewise, clinicians already had a tool to distinguish MLL from ALL--a translocation from chromosome 11 fused onto other chromosomes, discovered by Stanley Korsmeyer and others in the 1970s. Yet the new gene expression study tells oncologists something fresh and potentially important. Compared to ALL, RNA activity in MLL is silent or underactive in 1,000 genes and overactive in about 200. One gene really stands out. The gene flt-3 codes for a tyrosine kinase that fuels cell growth and is mutated in many cases of AML, which shares some gene activity patterns with MLL. Molecules that inhibit flt-3 are in preclinical development. A similar finding in another cancer, chronic myelogenous leukemia, led to the drug Gleevec, which blocks the overactive enzyme. Armstrong is developing a mouse model of MLL to test new approaches to treatment. Contrast this to the lung cancer study, in which Meyerson and his colleagues cannot be certain if their four new adenocarcinoma subtypes have much clinical meaning--in part because blood diseases are decades ahead of lung tumors in terms of scientific study and understanding. Finding biological meaning in the data is just one hurdle. Another problem is figuring out whether old statistical methods work with the new databases. It's not going to get easier; the multiclass study suggests more than a handful of genes will be needed to diagnose cancers by gene expression. Statistically, it is like turning the Nurses' Health Study on its side, Golub said. There, researchers analyze a modest number of variables in tens of thousands of people. Gene expression studies, on the other hand, analyze tens of thousands of variables (genes) in a modest number of tumor samples. "It's a nightmare," Golub said. "That alone makes most statisticians run out of the room screaming." Golub characterizes this group of studies as "small, tantalizing, encouraging preliminary studies. I wouldn't want to give the impression that these are slam dunks," he said. "We are starting to see how there may be some applications that may have clinical value, and we're starting to think about how to deliver that. External validation is essential, particularly things that impact on the standard of care for patients." --Carol Cruzan Morton |
