DELFI approach as ‘pretest’ in early cancer detection

A person’s cfDNA is derived largely from white blood cells, “and when you do whole genome sequencing of that DNA, essentially scooping up all the DNA that’s in the blood and looking at size and location, you can obtain a fragmentation profile,” Dr. Velculescu said.

In individuals with cancer, the fragmentation profiles are typically different, he said, because the tumor-derived cells have changes in the way their chromatin has been organized. “It’s no longer packaged in the same way in an ordered fashion. In fact, the differences in the size of the nucleosomes and the nucleosomal DNA that’s wrapping the nucleosomes, as well as the distance between them, the open and closed regions of the genome, and other genomic characteristics can end up affecting this fragmentation profile.”

Dr. Velculescu and his colleagues used machine-learning approaches to compare these different profiles and distinguish those with cancer from those who are healthy. And because the residual fragmentation profile still has information on the tissue of origin, he said, it can be used to look at the source of the tumor-derived DNA.

“When you think of mutations, you typically think you have one to hundreds of mutations in a targeted panel,” he said. “When you’re thinking of methylation, perhaps you can get thousands of these changes that one can evaluate in a targeted way. And when you’re looking at all the fragmentation differences, there are potentially millions of differences to identify.”

This can greatly expand the sensitivity of such an approach, he said. “It increases the number of shots on goal and can identify not only those individuals who have cancer but also the tissue of origin.”

In a pilot analysis, his team isolated cfDNA from approximately 4 mL of plasma from 30 healthy individuals and eight patients with stage I to III resectable lung cancer and performed whole genome sequencing at approximately 9× coverage. The fragmentation profiles of the 30 healthy individuals had similar and consistent patterns, Dr. Velculescu said, while fragmentation profiles of the eight lung cancer patients had dramatic changes, “in many cases occurring at multiple regions throughout the genome.”

“We wondered whether one can evaluate this at lower coverage” of the genome than 9×, he said. His team subsampled whole genome sequencing data and determined that altered fragmentation profiles from cancer patients were identified as low as .5× coverage. The benefit is that this approach is now broadly applicable and inexpensive, he said.

The questions they had initially, when using this approach, were what is the source of the healthy cfDNA, why have a profile at all, and why isn’t the DNA uniformly distributed throughout the genome. So they isolated the nuclei of WBCs, evaluated nucleosomal DNA from those individual cells, and sequenced them.

“When you sequence that nucleosomal DNA coming from the cells in healthy individuals, you can see a profile,” he said, “and it turns out that profile is similar to that of healthy individuals and their cell-free DNA, demonstrating that the source of cell-free DNA in a healthy individual is nucleosomal DNA. And the profile we’re seeing genomewide is the same profile you see from these cells.”

What about those who have cancer? Cristiano, et al., wrote in their 2019 article, “In contrast to healthy cfDNA, patients with cancer had multiple distinct genomic differences with increases and decreases in fragment sizes at different regions.” While chromosomal gains and losses present in the tumor of one individual with cancer were easily detectable in the cfDNA, Dr. Velculescu said, “the profile genomewide is what makes the difference. If you just looked at the overall distribution of sizes, for example, of these cell-free DNA patterns, and looked in the tumor versus the healthy individuals,” the overall cfDNA fragment size differences are small and would not have been useful for distinguishing cancer-derived cfDNA from healthy cfDNA. “It’s a genomewide pattern that is useful.”

Dr. Velculescu and his team then expanded the study, performing WGS at 1× to 2× coverage of cfDNA from 215 healthy individuals and 208 patients with various (largely early stage) cancers: breast (54), colorectal (27), lung (12), ovarian (28), pancreatic (34), gastric (27), and bile duct (26). “Of course, one can use machine learning to take advantage of all this information,” he said. The team implemented a gradient-tree boosting, machine-learning model to examine whether cfDNA can be categorized as having characteristics of a patient with cancer or a healthy individual.

“This allows us to consider this not necessarily a multianalyte test but a multi-feature test that can take this multitude of information and utilize it to develop the best algorithm to, in a high-confidence way, detect and distinguish those individuals with cancer from those who are healthy.”

Ultimately, they were able to identify with high specificity between 60 and nearly 100 percent of individuals across the seven cancer types while detecting few abnormalities in those without cancer, Dr. Velculescu said.

“This led to an analysis of the sensitivity and specificity. An ROC curve for the overall cohort of cancer patients was 0.94 in the study, and turned out to be much higher than previously proposed methods,” such as looking at mitochondrial DNA in the blood or chromosomal copy numbers, he said.

Dr. Velculescu’s team was the first to propose a machine-learning approach for genomewide fragmentation patterns, he said, “and the performance of the DELFI approach was higher in this analysis than other approaches and shows highest performance across the different cancer types analyzed.”

One can consider combining the fragmentation profile approach with other types of alterations, he said. For some set of patients analyzed, his team also performed targeted deep sequencing, looking at specific mutations in the cfDNA. Of 126 individuals analyzed with the two approaches, “we can detect about 82 percent with DELFI and about 66 percent with mutations.” Looking at the DELFI score and mutations in combination led to the identification of about 91 percent of patients with cancer.

The study findings demonstrated that “genomewide fragmentation profiles are a universal feature of human cancer,” Dr. Velculescu said. “They can be used to identify those individuals with cancer and are successful” in late- and early-stage disease.

Identifying the source of the cancer is a remaining challenge. The fragmentation profiles were different among the different cancer types, Dr. Velculescu said, and he and colleagues tested whether the information could be used to identify the tissue of origin. They found they can predict the tissue of origin accurately about 61 percent of the time; about 75 percent of the time one of the top two predictions would be accurate, he said.

“This is an important first step,” Dr. Velculescu said. “We anticipate that with larger numbers of tumors that are analyzed for each of these tumor types, we will be able to improve these performance metrics.”

The DELFI method can have applications other than screening, Dr. Velculescu said in the recent interview. Monitoring of therapy is one example. Another is early recurrence detection. In the Mathios, et al., study of lung cancer patients published this year, Dr. Velculescu and coauthors looked after resection at whether these individuals had recurrence. “And we were able to identify in a number of them that they had recurrence of disease months prior to when they were ultimately diagnosed.”

Prognosis is yet another application. Individuals who had a higher DELFI score had a worse outcome. This was true even after accounting in a multivariate analysis for such things as the stage, size, and type of cancer and the histology, Dr. Velculescu says. One reason may be that the tumor is more aggressive, “because this fragmentation is a measure of the disorganization, in a way, of the cancer genome.” The second reason may be that the DELFI approach is detecting occult metastases.

Next up, Dr. Velculescu says, is a 1,700-person first-of-a-kind national clinical trial—DELFI L101—sponsored by the Johns Hopkins University spin-out, Delfi Diagnostics, with U.S. participants who are healthy individuals, individuals with lung cancer, and individuals with other cancers. “Over time we envision the technology being applicable to other cancer types as well,” he says.

Cost-effectiveness is an important aspect of the DELFI approach, Dr. Velculescu says. “Sometimes technologies that are developed are quite expensive, and if we have tests that are too expensive, they end up being like VIP tests which then don’t help from a public health perspective. And early cancer screening and detection is a public health effort. So if one does such tests, and they’re only applicable to one percent of a population, then we’ve failed as a society to help those who need this most.”

Amy Carpenter Aquino is CAP TODAY senior editor.