NEW YORK – Cancer early detection firm Delfi Diagnostics is pushing forward with a newly enhanced platform, combining its original fragmentomic technology with a new machine learning method for gleaning cancer signals in the blood.
The new method, which analyzes genome-wide mutational patterns, boosted the performance of Delfi's assay above 90 percent in a recent lung cancer study.
Delfi licensed the new mutational profile strategy, called GEMINI (Genome-wide Mutational Incidence for Non-Invasive detection of cancer), from Johns Hopkins last month after the university researchers who founded the company showed that it added value to the firm's existing test, while still relying on circulating cell-free DNA.
In earlier studies, Delfi's technology achieved almost as high a detection rate but saw that drop significantly for the earliest stages of cancer. Combining mutational profiles with its existing algorithms for fragmentation patterns, the company has now pushed up both its overall performance and, crucially, its success in detecting stage I disease.
Victor Velculescu, Delfi's cofounder and CEO, said that when the company was initially spun out, mutations were viewed as a poor source for cancer early detection and screening, despite their value for liquid biopsy in other areas of precision medicine.
"Mutations were initially thought to be useful for liquid biopsies and they are in many cases, but they're also very, very difficult for screening for two reasons primarily," Velculescu said. "One is because you typically don't get enough of the genome analyzed. You're just looking at a targeted region, and many patients just don't have a mutation in that region. Or, even if they do have a mutation, with the number of molecules you look at there's typically a limited number in the blood so you may not have that mutation in that sample of blood."
Researchers have tried to solve this with ultra-deep sequencing, but that becomes expensive, Velculescu said, adding that "on top of that, they collect like five tubes of blood … and then it still doesn't work."
This signal-to-noise issue is linked to the nature of sequencing itself, and also to a process called clonal hematopoiesis, whereby white blood cells acquire somatic mutations, naturally over time. "We all get some small number of mutations in a lot of key genes as we age. They may not be affecting things, but they do create this noise [that] can be confused with cancer," Velculescu said.
When it launched, Delfi decided to avoid these pitfalls by forgoing mutations altogether, and looking instead at genome-wide patterns of DNA fragmentation that can mark the presence of cancer regardless of mutational status.
Velculescu said that he and his colleagues began developing GEMINI about a year ago. The resulting method, which uses machine learning to assess whether samples feature cancer-like mutation patterns, relies on the fact that in cancer different parts of the genome have both different types and different numbers of mutations.
"Many people didn't recognize this because typically they're looking at one gene of interest that is driving the cancer and aren't interested in so-called passenger mutations," Velculescu said.
Studies in tumor tissue had begun to demonstrate that there were broader genome-wide mutational frequency patterns that could distinguish cancer from normal cells. Delfi's task was to find a way to do the same in cell-free DNA and to do so without the need for deep sequencing.
The company has not been alone in that pursuit. Memorial Sloane Kettering's Luis Diaz last year published on a similar approach, which he and his team dubbed Pointy. According to Velculescu, the Hopkins method is unique due to its strategy for overcoming confounding background mutations by examining regional mutational frequency.
"What was the sort of novelty here in this approach is that one can come back and identify the regions of the genome which are more frequently mutated … and identify those regions that are important for cancer," Velculescu said. These regions, the company found, are not only specific to cancer versus normal cells and versus clonal hematopoiesis but also unique to different cancer types.
"What we were able to do is to identify these regions and then [use machine learning to] ask, for any new sample that comes in, whether the mutations … looked more like a cancer patient or a healthy individual," he said.
The "very convenient" aspect of this, he added, is that this information can be gleaned from the same data that Delfi already uses in its fragmentomic method. Assuming the two cancer signals — DNA fragmentation and genome-wide mutation patterns — offer independent value, the Delfi developers hoped that combining them would boost their ability to detect cancer.
In a paper published last month in Nature Genetics, Velculescu and colleagues described a variety of experiments developing and validating GEMINI, which culminated in them testing a combination of Delfi and GEMINI in 163 samples from lung cancer cases and non-cancer controls. They reported that the paired approach achieved an area under the receiver operator curve of 0.931. AUC is a common measure of diagnostic accuracy that considers both sensitivity and specificity. The closer to 1, the more accurate the test is.
Performance notably remained high even among cancers in their earliest stage. In a subset of patients with stage I lung cancers, GEMINI demonstrated an AUC of 0.81 alone and 0.86 when combined with Delfi.
The combined test also showed similarly high performance in other applications including therapy response monitoring and distinguishing subtypes of lung cancer.
"I think one of the beauties of this analysis and approach is that it shows just in general that [we] can add … different types of features that are not contained in the original fragmentation profiles … to increase the performance," Velculescu said.
Among other moves the company has made in this vein, Delfi has also experimented with boosting accuracy by examining chromatin changes and transcription factors.
"There's a great amount that we've learned and that I think we still have yet to learn from what I love to call the cell-free DNA universe," Velculescu added. "The benefit of that is that we can have what you can think of as a multi-feature test … [and] I think that is likely to be a lot more powerful than what people called a multi-analyte test where … you're improving performance, but you're driving up the cost of the test, as well."
In its push to the clinic, Delfi is currently in the midst of two lung cancer screening trials, though it plans eventually to develop its technology in other tumor types like liver cancer. One trial is a case control study, and the other, CASCADE, is a prospective effort. Velculescu said the company expects to report outcomes in the near future.