NEW YORK – Delfi Diagnostics has zeroed in on lung cancer as the first application for its liquid biopsy technology, a machine learning-driven method that analyzes patterns of cell-free DNA fragmentation to detect the presence of cancer in assumedly healthy individuals.
The firm and its academic collaborators published promising new data in Nature Communications last week, from a study of high-risk individuals undergoing low-dose CT scans to detect incipient lung cancer. Across more than 350 patients, Delfi's blood test picked up a tumor signal in a large proportion of those who were ultimately diagnosed with cancer, and this performance held up when applied to an independent validation set.
According to Delfi's chief medical officer, Peter Bach, the results were strong enough that the company has already begun a follow-up prospective case-control study enrolling at least 1,700 patients in order to lock in a classifier, with plans for a follow-up validation study that will then generate the necessary evidence for a commercial launch.
Targeting lung cancer made sense for the company on several fronts, according to Delfi founder and CEO Victor Velculescu.
For one, it represents the leading cause of cancer deaths, with the majority of tumors detected at late stages when prognosis is poor. In addition, the US Preventive Services Task Force already recommends screening, via low dose CT imaging, for high-risk individuals aged 50 years and over.
As a result, there isn’t a need to break new regulatory ground as is the case for companies like Grail, he said, who are advancing multi-cancer screening tests that don't fall under existing governmental guidelines.
Low dose CT has already proven effective in lowering lung cancer mortality, but reports have found that as few as 6 percent of eligible individuals get screened. Delfi is hoping that a blood test — less expensive and easier to perform in greater numbers of locations — could significantly improve these numbers.
"The benefit of a test like this is that you have a very simple, easy-to-use, widely accessible, cost-effective approach that could be broadly utilized and would increase the [number of] individuals that are prescreened and then make it through this particular pathway," Velculescu said.
In their Nature Communications study, Delfi researchers and their colleagues initially tested samples from 365 individuals screened and followed as part of a seven-year Danish study called LUCAS. The majority were at high risk for lung cancer and had smoking-related symptoms such as cough or difficulty breathing.
Delfi's only other published study of its fragmentomic method had applied it to a set of retrospective samples from several different cancer types. For the LUCAS cohort, the company started fresh, retraining a new algorithm specific to lung cancer in the prospectively screened cohort.
Overall, 129 individuals in the study were determined to have lung cancer after imaging and biopsy, while the remainder had histologically proven benign nodules or were not biopsied due to low clinical and radiographic suspicion for cancer.
When the Delfi team examined cell-free DNA fragmentation patterns across the cohort, they saw that profiles were highly consistent among non-cancer individuals, including those with nonmalignant lung nodules. "In contrast, cancer patients displayed widespread genome-wide variation," the authors wrote.
Calculating Delfi scores for the patients and assessing their discriminator performance, the group concluded that they could distinguish cancers from non-cancers with high sensitivity and specificity, represented by an area under the curve of 0.90. Detection of stage I disease was more difficult, with the AUC dropping to 0.76, but the statistical analyses showed that AUCs remained high — between 0.89 and 0.92 percent — for stages II through IV.
To externally validate the predictive performance of Delfi, the investigators then locked in a fragmentation score cutoff point based on the LUCAS cohort and applied it to an independent set of blood samples representing 385 lung cancers and 46 controls. According to the authors, the performance remained similar, with sensitivities from about 70 percent to 100 percent depending on stage.
According to Bach and Velculescu, the practical implementation of Delfi in the clinic would be as a prescreen, with positive patients referred for imaging. Based on the performance of the method in the LUCAS cohort, the investigators calculated that if the test had been implemented for these individuals, the combined blood test and imaging would have led to detection of 90 percent of lung cancers, including 80 percent of stage I cancers, and would have reduced the number of LDCT-induced unnecessary procedures by 50 percent.
"The idea is that this would be beneficial because you get more people into the system that are being screened and a higher fraction of individuals that are image positive actually end up having cancer, so the whole approach is actually one that gets more people detected with cancer, and in a more cost effective manner for the whole health care system," Velculescu said.
Important for this population, Bach added, the study also showed that the Delfi fragmentation method isn't confounded by noncancerous lung conditions like benign nodules and respiratory disease. For a high-risk population, which is already being predefined based on things like age and behavior, the demands on a screening test are higher than in a broad-based population, he said.
"My understanding from the literature is that the vast majority of lung cancer biomarkers have been challenged by exactly this issue. They are oftentimes altered in inflammatory conditions and in individuals with nodules, but without cancer. … So that's one of the benefits of this approach, that … it wasn't affected by other conditions. It wasn't affected by smoking. It wasn't affected by inflammation, or by COPD," Velculescu added.
Moving forward, Delfi is taking a two-step strategy for test development. The first is the aforementioned case-control study, which Bach said will recruit up to 2,000 subjects.
This will be used to finalize a clinical classifier, which the firm will then validate in a prospective study of individuals in its intended use high-risk population. "We'll draw their blood and assess what the test says about the presence or absence of lung cancer and how that is predictive of what's then seen on their CT scans," Bach said.