By David N. Leff
Imagine a tribunal that handed down a sealed verdict against a large group of defendants. It sentenced some of the accused to death, and acquitted others as not guilty. But this secret decision didn¿t name who would die and who go free.
That is precisely the sword of Damocles poised over the heads of people diagnosed with a cancer of one blood-cell type, called diffuse large B-cell lymphoma (DLBCL). It¿s the most common lymphoid malignancy in adults ¿ and curable in fewer than half of patients diagnosed with DLBCL. The treatment of choice for this disease is a four-drug cocktail of chemotherapeutic compounds called CHOP ¿ cyclophasphamide, adriamycin, vincristine and prednisone.
¿In almost all cases, there is an initial response to this standard chemotherapy,¿ observed lymphoma researcher and clinician Margaret Shipp, at the Harvard-affiliated Dana-Farber Cancer Institute in Boston. ¿For patients whom it cures of their disease, tumors shrink and shrink and eventually go away.¿
CHOP chemotherapy can actually reach malignant DLBCL cells all over the body, so when it¿s effective it will kill those malignant cells. The tumor will die and be dissolved by the body¿s scavenger mechanism. When that cure is not effective, the tumors regrow and eventually cause a variety of symptoms that ultimately result in death.
Shipp, who directs Dana-Farber¿s lymphoma program, is lead author of an article in the January 2002 issue of Nature Medicine, titled: ¿Diffuse large B-cell lymphoma outcome prediction by gene-expression profiling and supervised machine learning.¿ Its senior author is hematologist Todd Golub, who holds a joint appointment at Dana-Farber in pediatric oncology and the Whitehead Institute/MIT in Cambridge, Mass., in gene expression profiling.
¿From a clinical perspective, people have been puzzling for many years,¿ Shipp told BioWorld Today, ¿that a significant percentage of DLBCL patients can be cured of their disease with combination chemotherapy. But a large number are not successfully treated with that same type of therapy, and go on to die of the disease. So by definition, that tells us that there is likely to be some intrinsic heterogeneity among the different lymphomas that we can¿t see at a clinical level, which is influencing their response to treatment.¿
Making Invisible Outcome Molecules Visible
¿So several years ago,¿ she continued, ¿we developed a clinical model that identified people who were more or less likely to be cured with available chemotherapy. But it didn¿t address the inherent questions that account for the cellular and molecular differences among subsets of tumors that we couldn¿t identify as different under the microscope.
¿So we turned to microarray gene expression profiling,¿ Shipp recounted, ¿which allowed us to take tumors that look for all the world the same under the microscope, seek out their molecular signature, and actually predict their clinical outcome.¿
She and her co-authors took a series of tumors excised from 58 DLBCL patients for whom they had the complete clinical picture ¿ knowing the groups that were cured of their disease with chemotherapy, and those that were not. They then asked: ¿Could we use gene expression profiling to identify molecular correlates and outcomes?¿
¿The 58 patients,¿ Shipp pointed out, ¿were not preselected to include in this pilot study. They are a group for whom we have the frozen tumors available to be analyzed, and also their clinical information available to be able to determine the extent to which the gene profiles and medical data related to each other. Without both parts, obviously, we couldn¿t do clinically relevant profiling.
¿A number of those donors are still alive,¿ she observed, ¿and a number have died of their disease. Those who had curable tumors went on to live their lives, and DLBCL became part of their history.
¿The platform that we used,¿ she went on, ¿was the Affymetrix [Inc.] oligonucleotide microarray chip, loaded with probes to 6,817 genes. These represented a sizeable fraction of the human genome, that can be analyzed in the 58 tumor tissues we tested.
¿The method we used to analyze this raw data,¿ Shipp recalled, ¿was the supervised machine learning approach. What that means,¿ she explained, ¿was that we had a hypothesis: Two groups of DLBCL that differed with respect to one fundamental property ¿ their curability with respect to standard chemotherapy. We took the data and used different supervised learning algorithms to see if we could identify a gene expression signature that is associated with that difference. Then when we identified a very large signature relevant to that property, we pared it down to a smaller signature that would really capture the fundamental difference, but with a much smaller set of genes.
¿With that prediction algorithm,¿ Shipp went on, ¿we were able to identify 13 key genes, by which we could capture a significant component of the differences in curability of the tumors.¿
First Clinical Trial In Spring Of ¿02
¿We are already in the process right now of credentialing certain of these genes as targets for rational therapy and looking at ways we can inhibit the gene products in those pathways, in order to incorporate them into treatment approaches for this disease. And we are using the same approach on a much larger expanded series of tumors, attempting to validate our initial findings, and confirm them by protein expression as well.¿
Shipp described her tentative strategy for targeting therapeutic genes and their protein products:
¿Different genes require different means,¿ she commented. ¿For instance, a couple of gene products in our first signature are enzymes that are critical for regulating specific pathways. There are already in a couple of cases inhibitors of those enzymes that have made it fairly far along in clinical development for patients. So we plan to take one or more of those enzyme inhibitors and ask whether these might block the growth of DLCBL-cell lymphomas. If it did, we¿d think about incorporating it into a clinical treatment trial.
¿We have one clinical study,¿ she added, ¿that we anticipate opening up in the spring of 2002, based on one of the targets in our initial gene expression profiling.¿
Meanwhile, Golub, the paper¿s senior author, ¿has made progress extending the team¿s technique to other malignancies. He and his group,¿ Shipp concluded, ¿have been looking at heterogeneity in breast tumors and lung cancer, with ongoing work in prostate cancer and melanoma.¿