The many companies developing liquid biopsies to try to detect cancer early have so far largely mined the blood in search of things like mutations and epigenetic changes in human DNA shed by tumor cells.
Now, new research raises the possibility that liquid biopsies could be used to spot cancer in a totally different way: by hunting for the DNA of bacteria and viruses released from tumors into the bloodstream. It’s a hypothesis that, if validated with more study, could usher in an entirely new class of diagnostics for cancer.
In a study published Wednesday in the journal Nature, a team led by researchers at the University of California, San Diego, reported that they have developed machine learning models that, in early-stage testing, could identify and distinguish between different types of cancer based on microbial signatures in the blood.
advertisement
The senior author of the study, the leading microbiome researcher Rob Knight, called the research “one of the most significant things to come out of our lab” since he moved to UCSD five years ago.
“It’s introducing a completely new kind of information that you can get out of a liquid biopsy — where we would expect that new information would allow us to see things that are missed by techniques that just focus on the human DNA,” Knight said.
Knight and two of his co-authors on the study have founded a San Diego-based startup, called Micronoma, to try to commercialize a cancer test based on the research.
Doing so won’t be easy.
Geoffrey Oxnard, a medical oncologist at Dana-Farber Cancer Institute in Boston, praised the new study for putting forth a “fascinating” and “relatively underleveraged” idea. But he also questioned whether the method will be able to be used as a cancer detection test.
“I would want to see its performance in large numbers of controls with all sorts of other inflammatory or infectious conditions because I would worry that it could be vulnerable to false positives when upscaled to larger populations,” Oxnard said.
Oxnard is one of the lead investigators for a longitudinal study, sponsored by the liquid biopsy company Grail, that’s trying to map the genomic cancer signals in the blood of people with and without cancer. Grail is one of the liquid biopsy developers interested in human DNA; for its cancer test, it’s homed in on methylation, an epigenetic change across the whole genome.
To develop their new method, Knight and his collaborators began by looking at sequencing data from more than 18,000 tumor samples covering 33 different types of cancer from The Cancer Genome Atlas, a large repository maintained by the National Institutes of Health. The Knight lab focused on the reads mapped to microbes, which are usually thrown out by researchers interested in human DNA, and used the data to train machine learning models to distinguish between different cancer types. (The atlas data weren’t collected with the goal of being mined this way, a limitation of the research.)
Knight and his team then set out to test their models. They collected blood samples at UCSD from 59 patients with prostate cancer, 25 patients with lung cancer, 16 patients with melanoma, and 69 healthy volunteers who did not have cancer or HIV. For each of the samples, they sequenced the microbial signatures and then tasked their models with analyzing those data.
The models performed well. They correctly identified 86% of the people who actually had lung cancer and 100% of the people who did not have lung cancer. And they correctly distinguished between cases of prostate cancer and lung cancer 81% of the time.
Those results come with caveats. As with any machine learning experiment in this context, there’s a risk that the training data has been modeled too well, a limitation known as overfitting, said Curtis Huttenhower, a researcher at the Harvard T.H. Chan School of Public Health who studies the role of the microbiome in health and disease. (Huttenhower’s lab has collaborated with Knight’s lab on several projects.)
“You have a noisy high-dimensional measurement in a complex structured population where, even as they show in the paper, there are lots of technical things that can drive differences. And they’ve done a good job of minimizing many of those,” Huttenhower said. “But it’s almost impossible in this sort of secondary post-hoc analysis to say: ‘Yes, we’ve eliminated all of the possible technical things that a machine learner could overfit to.’ And you can kind of see that reflected in the very, very, very high accuracy they achieved in the paper.”
The new study builds on an emerging body of research showing that tumors are not sterile environments but rather serve as habitats for bacteria and viruses. It’s not clear where the microbes come from — it may be the mouth or the gut or another site of origin — but once they get to the tumor, they live and grow there and secrete their DNA into the bloodstream.
It’s not clear what role, if any, these microbes are playing in the growth of the cancer. It’s a classic the-chicken-or-the-egg problem, said MD Anderson Cancer Center’s Jennifer Wargo. As she puts it: Are these microbes “actually contributing to the cancer development — and therefore drivers? Or are they just merely passengers?”
Wargo, a surgical oncologist who leads a program at MD Anderson to profile the microbiome to understand its impact on cancer patients and healthy individuals, called the Knight lab’s study “incredibly provocative” and “a call to action” for more research in the field.
The Knight lab’s study was supported with funding from the NIH, as well as a UCSD fund for microbiome research.
While the new study is to date the most ambitious and wide-ranging work of its kind, it is not the first to demonstrate the principle that microbial signatures in the blood can be used to detect cancer. In a 2017 study published in the New England Journal of Medicine, a team of researchers in Hong Kong reported that they had developed a method to probe the blood for DNA of the Epstein-Barr virus. The idea was to screen for a type of head and neck cancer known as nasopharyngeal carcinoma, which is prevalent in Southeast Asia and has been linked to EBV.
