A metabolite-based diagnostic developed by researchers at the Georgia Institute of Technology identified women with ovarian cancer with 100 percent accuracy in a recently published 94-subject trial.
Although additional, larger trials are required to validate the test, the initial results suggest it could be useful in screening high-risk patients for ovarian cancer and possibly even as a screening diagnostic for the general population, John McDonald, chief research scientist at Atlanta's Ovarian Cancer Institute and one of the paper's authors, told ProteoMonitor.
The fifth leading cause of cancer-related deaths among women, patients diagnosed in the disease's early stages have a greater than 90 percent five-year survival rate. Because early-stage patients are typically asymptomatic, however, ovarian cancer often goes undetected until its later stages. The five-year survival rate for these late-stage patients is roughly 20 percent.
There is need, therefore, for diagnostics capable of detecting ovarian cancer in its early stages, and the disease has been a primary focus of diagnostic development, with a variety of protein biomarker-based tests – including Vermillion's OVA1, Correlogic's OvaCheck, Healthlinx's OvPlex, and Arrayit's OvaDx – in some stage of commercialization.
According to McDonald, the prevalence of such protein-based ovarian cancer diagnostics was one of the main reasons the Georgia Tech group decided to look at the metabolome instead of the proteome for potential biomarkers.
"We wanted to get into something that was less crowded," he said.
He added that another potential advantage of the metabolomics approach is easier mass spectrometry analysis.
"Because the metabolome is much smaller than the proteome for mass spec analysis things are spread out better, so we get better resolution," he said.
The Georgia Tech test, which is described in the current issue of Cancer Epidemiology, Biomarkers, & Prevention Research, used direct-analysis-in-real-time mass spectrometry to measure thousands of metabolites in subjects' blood samples and then classified them with a functional support vector machine-based machine-learning algorithm.
The researchers evaluated the method via two different approaches. In the first – a 64-30 split validation – they used 64 of the subjects as an algorithm training set and the remaining 30 as an independent test set. Using this technique they achieved 100 sensitivity and 100 specificity.
In the second approach – leave-one-out-cross-validation, or LOOCV – the researchers used 93 of the samples as the training set and the remaining sample as the unknown, running the samples sequentially until each of the 94 had been measured as an unknown. With this method they classified all but one of the 94 samples correctly, for an overall accuracy of 98.9 percent – 100 percent sensitivity and 98 percent specificity.
McDonald was "quite shocked," by the results, he said. "I was expecting the metabolome not to be a good way to go because I thought metabolites would be highly variable. I was really surprised that we were able to get such accuracy out of it."
One thing contributing to the test's accuracy, McDonald said, is the large number of markers it employs.
"In ovarian cancer, the single protein that's commonly used [as a biomarker], CA-125, is not a very accurate test," he said. "The reason for that is that all cancers are variable. So if you're relying on a single biomarker, it's very unlikely that that single biomarker will be 100 percent accurate or even 99 or 95 percent accurate."
"Even going from one to five [biomarkers] increases accuracy tremendously. In our case we're using at the minimum 2,000 to 3,000 features. That should in theory give us an even higher degree of accuracy," he said.
By comparison, most protein-based tests that are commercially available or under development use a handful of markers. Vermillion's OVA1, for example, analyzes five protein markers, including CA-125. The HealthLinx OvPlex test also uses five proteins, including CA-125, and the company is currently evaluating two additional markers to add to the test (PM 6/18/2010).
Because of the low prevalence of ovarian cancer, screening tests need to be highly accurate to avoid a large number of false positive results. In fact, McDonald said, for a screening test to be useful for the general population it would need to demonstrate 100 percent accuracy.
The researchers used DART-TOF MS on a JEOL AccuTOF instrument to isolate the thousands of metabolomic features used in the test. Initially, they prepared the samples using conventional liquid chromatography, but realized that with DART they could get sufficient resolution without any separation steps.
"The problem with [liquid chromatography] is the time involved," McDonald said. "With the DART technique you don't have to do any sample preparation. You just put the sample in and it vaporizes and goes in. We know we lost some resolution going to [DART], but we're still getting 20,000 features, so we're well above what we need for diagnostic purposes."
[ pagebreak ]
McDonald's lab is also running pancreatic and lung cancer samples on the platform.
"Our technology isn't limited to ovarian cancer," he said. "Ideally, in the future, the models would be stored on a computer and as a patient you would come in with a single drop of blood, get your profile run against a database and you could in theory be tested for multiple diseases in a single assay."
Looking for Validation
At present, though, McDonald's team is still working to validate the ovarian cancer test. The researchers are applying it to an additional set of 500 ovarian patient samples and expect to have data from that study in around six months.
"We won't know if the accuracy holds up the way that we think it will until we run many more samples," McDonald said.
The team is also trying to assemble samples from early-stage patients for testing. Such samples are difficult to find, however, given that most early-stage patients show no signs of the disease. McDonald added that the researchers planned to secure them by establishing partnerships with medical centers around the country.
The researchers' approach is "very interesting," Correlogic CEO Peter Levine, who was not involved in the study, told ProteoMonitor. Like McDonald, however, he said it would be hard to judge the diagnostic's usefulness until it had been tested on a larger number of samples.
"What they've done is they've trained, and they've tested. They haven't done a third validation. [Now] you need to take that algorithm and ... expose it to samples which were not used in training or testing," he said.
Levine added that the large number of markers in the test could lead to statistical overfitting, which can occur when a model has too many variables compared to the number of samples used.
He also raised concerns about its prospects for US Food and Drug Administration approval, noting that a diagnostic based on thousands of unknown metabolite markers was perhaps too far "from the FDA's comfort zone."
When developing its OvaCheck ovarian cancer diagnostic, Correlogic "moved away from the anonymous marker approach and went toward identifying specific analytes exactly out of that [regulatory] concern," Levine said.
Correlogic, which filed for Chapter 11 bankruptcy protection in July, has been seeking FDA approval for OvaCheck since 2004 and is currently involved in the second arm of its original clinical trial (GWDN 07/23/2010).
In addition to potential regulatory issues, the test has presented some interesting intellectual property complications, McDonald said. Specifically – how do you patent a test based on unknown analytes?
"We can't go in and patent specific metabolites, because in most cases we don't actually know what these features are," he said. "It's an issue I think the patent office hasn't really dealt with before. It may end up that you have to patent specific features for each disease you look at."