With Prognosis Device, Researchers Limit Unnecessary Therapy

If UC Berkeley biostaticians are successful, doctors may be able to tailor medical treatment and subsequent observation to prevent recurring tumors, limiting unnecessary therapy and expense.

There are roughly 100,000 new cases of colon cancer each year in the United States. Doctors often disagree about how best to treat these patients. Sandrine Dudoit, a professor at UC Berkeley's Division of Biostatistics and Alain Barrier, a surgeon from Paris, have collaborated to try to figure out which cases would most benefit from subsequent treatment.

"Patients with a similar stage of disease and a similar treatment may have radically different outcomes," Barrier said. "Being able to identify subgroups of patients at high risks and at low risks of recurrence would permit (doctors) to adapt the postoperative treatment ... to the individual case."

When colon cancer is detected, the tumor is surgically removed. Thereafter, however, there is inconclusive evidence as to how well chemotherapy will prevent recurrence. Since chemotherapy has significant side effects and costs, doctors try to limit the use of chemotherapy.

"The aim of our research was to assess the possibility to build a prognosis predictor, based on gene expression profiles, for non-metastasized colon cancers," Barrier said.

In a recent paper in the cancer journal, Oncogene, Barrier and colleagues presented this potential microarray technology-based predictor.

Microarrays are an increasingly common biological tool allowing scientists to look at all of the RNA content of a cell and determine the relative gene expression profile. Gene expression refers to which genes are turned on or off.

Barrier and his colleagues took surgical samples from past colon-cancer patients and analyzed their RNA. Because they used patients who had had their tumors removed at least five years ago, they could compare the gene expression of tumor cells from patients who had later had a recurrence with the gene expression from those who did not.

While the tumors may have looked the same, they found the gene expression differed. The gene expression differed in consistent ways, meaning the researchers could predict whether a sample was from a patient who would go on to have a recurrence or one who would not, with relatively good, though not perfect, accuracy.

Perhaps even more surprising, the researchers also found they could predict with equal accuracy, based on the gene expression in non-cancer cells from another part of the colon. In other words, seemingly healthy tissue from at least five centimeters away from the tumor had invisible changes that allowed the researchers to predict whether that person would experience recurrence.

"This result seriously questions the role of the tissue adjacent to the tumor, which is generally considered as normal," Barrier said. "It may be hypothesized that the whole colon is affected by some changes."

In order to get to this point however, Barrier and his colleagues had to establish a statistical model with which to analyze the gene-expression data. For this Barrier turned to Dudoit, an expert in the field of statistical analysis of microarray data. Barrier spent about a year in Berkeley working with Dudoit and colleagues who helped establish the approach Barrier refers to as "double cross-validation."

"Cross validation allows us to select the predictor and then evaluate its accuracy," Dudoit explains.

Essentially the researchers took groups of patient samples with known outcomes and used them to determine what subset of genes were most predictive and then used the other groups to test whether they predicted correctly, and then repeated this with other samples as the starting point.

This careful statistical approach was particularly necessary as they only had 18 patient samples with which to work.

The authors have since reproduced these findings with a larger sample size of about 100 patients, and those results are under review.

"The possibility of developing a prognosis predictor is certainly the most important finding of our work, since it might change the postoperative treatment and follow-up of colon cancer patients," Barrier said.

However, practical use of such a predictor is still some ways off. Microarray technology is still quite expensive, and even in order to conduct the larger scale studies to verify this approach, the scientists have run into difficulties. Finding tissue samples from patients who have at least 5 years of follow up, and which were also collected in an equivalent way, has limited their availability of samples.

Researchers hope that doctors can one day tailor treatment and observation from analyzing the gene expression of a patient's removed tumor.

Tags: