An international research team headed by scientists at the University of California, San Diego (UCSD), has developed a novel diagnostic approach that harnesses machine learning (ML) models to identify who has cancer, and often which type, by analyzing patterns of microbial DNA—bacterial and viral—in a single blood sample. The researchers trained and tested their machine learning models on the distinct microbial signatures identified in more than 10,000 tumor samples from patients with 33 tumor types. They then showed that the models were able to identify whether an individual did, or did not have one of three types of cancer, and differentiate between the three, using only the microbial patterns in their blood.
The researchers suggest that this newly discovered cancer-associated blood microbiome may have applications beyond cancer diagnostics. “This new understanding of the way microbial populations shift with cancer could open a completely new therapeutic avenue,” suggested Sandrine Miller-Montgomery, PhD, professor of practice in the Jacobs School of Engineering and executive director of the Center for Microbiome Innovation at UCSD, who is co-author of the team’s published paper in Nature. “We now know the microbes are there, but what are they doing? And could we manipulate or mimic these microbes to treat cancer?”
The reported studies have been headed by Gregory Poore, an MD/PhD student at the UCSD School of Medicine. Poore is conducting his graduate thesis work in the lab of Rob Knight, PhD, professor and director of the Center for Microbiome Innovation. Poore, Knight, and colleagues report on their findings in a paper titled, “Microbiome analyses of blood and tissues suggest cancer diagnostic approach.”
Cancer is traditionally considered a disease of the human genome, the researchers wrote, but more recent studies have suggested that the microbiome may also be involved. “ … recent studies have shown that the microbiome makes substantial contributions to some types of cancer; in particular, contributions of the fecal microbiome to gastrointestinal cancers.” However, they continued, the extent and diagnostic implications of any microbial contribution to different types of cancer aren’t known. And it’s a difficult field of study, because of the potential for sample contamination during collection, processing, and sequencing. Nevertheless, as the investigators noted, technologies to reduce this contamination risk are being developed. “The use of recently developed tools to minimize the contributions of contaminants to microbial signatures could enable the rational development of microbiome diagnostics.”
Poore’s interest in the link between microbiome and cancer was sparked in 2017 when he saw details of a study in Science that showed how microbes invaded a majority of pancreatic cancers and were able to break down the main chemotherapy drug given to patients. Poore’s family also had experience of pancreatic cancer. His grandmother, who was otherwise seemingly healthy, had been diagnosed with late-stage pancreatic cancer in one month, and died the next. “She had virtually no warning signs or symptoms,” he said. “No one could say why her cancer wasn’t detected earlier or why it was resistant to the treatment they tried.”
Poore became intrigued by the idea that bacteria and viruses might play a bigger role in cancer than had previously been considered. “Almost all previous cancer research efforts have assumed tumors are sterile environments, and ignored the complex interplay human cancer cells may have with the bacteria, viruses, and other microbes that live in and on our bodies,” commented Knight. “The number of microbial genes in our bodies vastly outnumbers the number of human genes, so it shouldn’t be surprising that they give us important clues to our health.”
For their reported studies, the researchers first looked at microbial data available from The Cancer Genome Atlas (TCGA), a database of the National Cancer Institute containing genomic and other information from thousands of patient tumors. To the team’s knowledge, it was the largest effort ever undertaken to identify microbial DNA in human sequencing data. From 18,116 tumor samples, representing 10,481 patients with 33 different cancer types, the team identified distinct microbial signatures associated with specific cancer types. Some were expected, such as the association between human papillomavirus (HPV) and cervical, head and neck cancers, and the association between Fusobacterium species and gastrointestinal cancers. But the team also identified previously unknown microbial signatures that strongly discriminated between cancer types. The presence of Faecalibacterium species, for example, distinguished colon cancer from other cancers.
The team used the microbiome profiles of these thousands of cancer samples to train hundreds of machine learning models to associate certain microbial patterns with the presence of specific cancers. “Using normalized data, we trained stochastic gradient-boosting ML models to discriminate between and within types and stages of cancer,” the scientists wrote. Tests showed that the resulting machine learning models were able to identify a patient’s cancer type using only the microbial data from their blood.
When the researchers removed the high-grade (stage III and IV) cancers from the dataset, they found that the blood-derived microbial data could distinguish earlier-stage cancers. The results held up even when the team performed the most stringent bioinformatics decontamination on the samples, which removed the majority of microbial data. “These results often remain valid even after extensive internal validation checks and decontamination, which at times discards more than 90% of the total data,” they wrote.
Knight, Poore, and the team then designed a study to test their models in a clinical setting. “To demonstrate the real-world utility of these results while benchmarking against plasma-based ctDNA assays, we evaluated the use of plasma-derived, cell-free mbDNA [blood-based microbial DNA] signatures to discriminate among healthy individuals and multiple types of cancer in a validation study while implementing gold-standard microbiology controls for low biomass studies,” they wrote.
To do this, the team analyzed blood-derived plasma samples from 59 consenting patients with prostate cancer, 25 with lung cancer, and 16 with melanoma, provided by collaborators at Moores Cancer Center at UCSD Health. Employing new tools they developed to minimize contamination, the researchers developed a readout of microbial signatures for each cancer patient sample and compared them to each other and to plasma samples from 69 healthy, HIV-negative volunteers, provided by the HIV Neurobehavioral Research Center at UCSD School of Medicine.
The results showed that the team’s machine learning models were able to distinguish most people with cancer from the controls. For example, the models could correctly identify a person with lung cancer with 86% sensitivity and a person without lung disease with 100% specificity. They could also often distinguish which participants had which of the three cancer types. For example, the models correctly distinguished between a person with prostate cancer and a person with lung cancer with 81% sensitivity. “These microbial profiles appear to discriminate within and between most types of cancer, including when using blood-based mbDNA at low-grade tumor stages and in patients without any detectable genomic alterations on commercial ctDNA assays,” the team noted.
“The ability, in a single tube of blood, to have a comprehensive profile of the tumor’s DNA (nature) as well as the DNA of the patient’s microbiota (nurture), so to speak, is an important step forward in better understanding host-environment interactions in cancer,” said co-author Sandip Pravin Patel, MD, a medical oncologist and co-leader of experimental therapeutics at Moores Cancer Center at UCSD Health.
“With this approach, there is the potential to monitor these changes over time, not only as a diagnostic, but for long-term therapeutic monitoring. This could have major implications for the care of cancer patients, and in the early detection of cancer, if these results continue to hold up in further testing.”
As Patel noted, diagnosis of most cancers is invasive, time-consuming, and costly, as it commonly requires surgical biopsy or removal of a sample from the suspected cancer site, and analysis to look for molecular markers. Several companies are now developing liquid biopsy methods that might diagnose specific cancers in a simple blood draw, and technologies that allow them to detect cancer-specific human gene mutations in circulating DNA shed by tumors. This approach can already be used to monitor progression of tumors for some types of diagnosed cancers, but is not yet FDA-approved for diagnostic use.
According to the researchers, one advantage of cancer detection based on microbial DNA, compared to circulating human tumor DNA, is its diversity among different body sites. Rather than relying on rare human DNA changes, the study results indicate that blood-based microbial DNA readouts may accurately detect the presence and type of cancers at earlier stages than current liquid biopsy tests, including cancers for which there are no known genetic mutations that can be detected by existing platforms. “Together, the results imply that microbial communities are unique to each cancer type and that our approach of normalization and model training to distinguish cancers based on microbial profiles alone can be applied more broadly,” they wrote in their published paper.
The researchers acknowledged that it’s still possible that blood-based microbial DNA readouts could miss signs of cancer and return a false-negative result. But they expect that their new approach will become more accurate as they refine their machine learning models with more data. And while false negatives may be less common with the microbial DNA approach, false positives—hearing you have cancer when you don’t—are still a risk.
The team also cautioned that even if a microbial readout indicates cancer, the patient would likely require additional tests to confirm the diagnosis, determine the stage of the tumor, and identify its exact location. Nevertheless, they wrote, “The high discriminatory performance among healthy control individuals and patients with multiple types of cancer using only cell-free mbDNA in plasma, while adopting more extensive internal and external contamination controls than TCGA, suggests that clinically relevant and retrospective testing using widely available samples would be feasible and generalizable.”
Knight also acknowledged that it will be a long road to translate these initial findings into an FDA-approved diagnostic test. Critical to this translation will be validation of the technology in much larger and more diverse patient populations. It will also be necessary to define what a “healthy” blood-based microbial readout might look like among diverse people.
The investigators want to determine whether the microbial signatures they can detect in human blood are coming from live microbes, dead microbes, or dead microbes that have burst open, dispersing their contents. This insight might help them to refine and improve their approach. As they concluded, “… many technical and biological factors limit the analysis of retrospective cancer sequencing data for low-biomass microorganisms, and advances in this field will require collaborations between cancer biologists and microbiologists. Nonetheless, our results suggest that a new class of microbiome-based cancer diagnostic tools may provide substantial future value to patients.”
The team hopes that their reported study may also help to direct new research in the field of cancer biology. “For example, it’s common practice for microbiologists to use many contamination controls in their experiments, but these have historically been rarely used in cancer studies,” Poore said. “We hope this study will encourage future cancer researchers to be ‘microbially conscious.'”
Knight and Poore have filed patent applications on their technology, which, together with Miller-Montgomery, they have spun out into a new company, Micronoma.
