By William Langbein
Since the 1990s, monitoring biomarkers in the blood has been one of few noninvasive tools that an oncologist can use to diagnose cancer. Reliance on single biomarkers has been driven by a chemistry approach to research and wet lab technologies such as 2D gel electrophoresis — one antigen or protein can be isolated, characterized, and clinically linked to the presence of an invading tumor or cancer.
But four years ago, FDA researcher Emanuel Petricoin teamed up with the chief of NCI’s pathology lab, Lance Liotta, to use a new mass spectrometry technology — namely Ciphergen’s surface-enhanced laser desorption (SELDI) ProteinChip — to re-examine the use of single biomarkers in cancer diagnosis.
Instead of painstakingly isolating disease specimens, SELDI could be used to map thousands of distinct proteins in hundreds of patient samples. The technology could generate a proteomic fingerprint in a matter of seconds.
Of course, mining that data required a higher-order analytical tool that wasn’t yet available. So, after a serendipitous introduction by their wives, Petricoin and the president of Correlogic Systems, a company in the business of developing pattern-recognition software for detecting Internet fraud, got together to create algorithms for analyzing the thousands of proteins mapped in a drop of serum. The tools would examine various subsets of distinct proteins and distinguish any reproducible differences in their patterns as expressed in normal and cancer cells.
If thousands of proteins could be detected and mapped simultaneously, Petricoin and Liotta reasoned, the pattern of the protein expression would be a richer source of diagnostic data than a single, unreliable marker circulating in the blood.
“If you think about it, looking for a single biomarker in the blood is a rather simplistic view of [detecting] how cancer works,” says Petricoin, who co-directs with Liotta the FDA-NCI Clinical Proteomics Program and is a senior investigator at FDA’s Center for Biologics Evaluation and Research. Even if a biomarker is legitimately associated with the presence of cancer cells, the marker could just as easily be produced by the loss of certain healthy cells, he notes.
To be sure, the same complaint could be lodged against the Liotta/Petricoin method. Patterns they detect aren’t definitive indicators of cancer, either. But their study of protein patterns in ovarian cancer, published in The Lancet a year ago this month, provided the oncology community with a more sophisticated tool and has since spurred new cancer pattern-recognition initiatives, including ones at NCI, Duke University, Johns Hopkins, and Eastern Virginia Medical School.
The Ciphergen system, says Petricoin, effectively opened the door to using mass spec in clinical diagnostics.
“We [took] a leap of faith that we could determine a proteomic bar code for cancer in serum,” he recalls. “Serum patterns in humans are mostly homogeneous, but certain differences [cause] change on a day-to-day basis, such as hormone levels, age, time of collection, and diet.” By mapping changes in serum protein patterns, Petricoin and Liotta expected to find cancer’s signature. The differences in the patterns themselves, not the individual proteins, would be diagnostic tools.
“Remember, mass spectrometry records the proteins as a pattern,” says Petricoin. “Whether the protein [levels] increase or decrease is irrelevant. It’s the overarching pattern that is important.”
The researchers acknowledge that the key to distinguishing between the patterns of proteins in healthy and disease tissues is the “training set,” which initially establishes the most accurate pattern present in the disease.
For instance, to establish the disease pattern in ovarian cancer, they assembled 50 blood samples from women at various stages of disease. Then they matched the base disease pattern against 50 samples from healthy women to distinguish the heterogeneity between the two sets. The biomarker typically used to detect ovarian cancer is a protein known as CA125. It is difficult to detect in early-stage patients, and often yields false positives because conditions such as endometriosis, menstruation, and pelvic infections also produce high CA125 levels.
After Correlogic conducted a series of iterative tests against the two sets, the company was able to establish a bedrock algorithm that could discriminate between malignant and benign samples. Correlogic further refined the algorithm’s heterogeneity definition by adding samples from 16 women with non-malignant conditions such as sinusitis and rheumatoid arthritis.
And then, a conspicuous breakthrough: In a blind test against samples of a separate group of 50 women with cancer and 66 with a non-malignant form of the disease, the protein pattern test correctly identified all 50 ovarian cases, including 18 with stage-one disease. Each of the spectra contained 15,200 data points — one for each protein or peptide in the sample. The defining pattern was a combination of five proteins that segregated the two sets with 100 percent accuracy.
The Liotta/Petricoin approach to using pattern recognition as a diagnostic is not without its detractors. For one thing, even though physicians widely use pattern-recognition tools to identify problems in other disease areas, such as electrocardiograms for diagnosing heart conditions, pattern recognition hasn’t caught on as a diagnostic tool in cancer.
Also, many people in the mass spec community challenge the low-resolution capability of Liotta and Petricoin’s early tests with the SELDI chip. While scientists agree that CA125 is an unreliable and late-appearing marker, it has been characterized and clinically coupled to ovarian cancer. In contrast, the protein pattern recognition algorithm detects the presence of unknown proteins that have not been linked to the tumor.
The two researchers concede that the success of pattern recognition depends largely on how well the training set is prepared, but they argue their early efforts are only scratching the surface of the method’s potential. Each training set learns and becomes more accurate as new samples are added. In addition, clinicians can pursue the function of the unknown proteins. If the function is irrelevant, the training set can be modified to narrow the differentiation of the pattern. “We’re not making a preconceived conclusion about what these proteins do,” says Liotta. “We’re looking at the pattern of change [in proteins] that correlates to the disease state.”
Snapshots From Chips
Aside from potential life-saving new options for cancer patients, Liotta and Petricoin’s work holds commercial promise for both Ciphergen and Correlogic.
In November, Correlogic signed an agreement with the nation’s two largest labs, Quest and LabCorp, to commercialize pattern-recognition tests that emerge from the company’s partnership with NCI.
And Ciphergen’s VP of business development Robert Mauer foresees a market for pattern diagnostics for complex diseases or basic tests for two to three markers that incorporate the predictive power of the pattern. For example, he points to a current lung cancer study that has revealed about 20 unique markers for that disease. Four of those 20 yield 95 percent of the information necessary for a diagnosis. It could be that only those four markers will be used as a test.
Liotta and Petricoin have expanded their initial study to retrospectively examine tissue and sera of another 500 to 800 ovarian cancer patients. Their trial should be concluded by July 2003 and Petricoin says he could be ready this year to submit a 510k filing to the FDA for a pattern diagnostic test for women at high risk for ovarian cancer and to distinguish between benign and malignant tumors.
The duo has also set its sights on developing a soup-to-nuts system of high-throughput mass spectrometry to establish protein pattern diagnostic tests for all types of diseases. “The key is the ability to develop reproducible sera from high-throughput mass spec,” says Petricoin. “We think the mass spec tools are available to run 100,000 samples. If you get the process down to [sample handling and reliable mass spec], diagnosing a disease could become modular.”
To that end, Petricoin says the company is evaluating new mass spec systems from ABI, Amersham, and others to determine which system can work most reliably in high-volume diagnostic labs that serve physicians.
Petricoin and Liotta’s tests won’t replace the need for biopsy procedures or a physician’s exam, but the patterns should be able to guide clinical actions and assure early intervention, especially in a disease like ovarian cancer where time is the enemy given the absence of reliable markers. Petricoin believes cancer advocacy groups will urge adoption of pattern- recognition tests even if it is only an adjunctive tool.
“This is a disruptive [diagnostic] technology made possible by mass spec,” acknowledges Petricoin, adding that he chose to have protein pattern recognition validated first by FDA review, rather than introducing it only to select reference labs. “It’s polarizing to the clinical chemists who have been measuring specific analytes for years and to some oncologists who have spent careers relying on biomarkers. It can be threatening to [previous] practice.” But as someone recently reassured Petricoin, if you’re disrupting the status quo, you’re probably on to something.