By Paul Nicolaus
January 29, 2019 | As a leading cause of death across the globe, cancer continues to claim lives and ring up cash registers along the way. The disease caused nearly 10 million deaths in 2018, according to the World Health Organization (WHO), and the estimated annual economic toll tops a trillion dollars. It’s no wonder scientists continue to search for clever ways to battle back.
Liquid biopsies are one approach gaining attention and traction thanks in part to their ability to test patients multiple times in less invasive ways than traditional approaches. As a tumor grows, dies, or responds to therapy, it sheds proteins and pieces of DNA. But it’s difficult to pick up on these faint cancer signals, which is why some are now looking to boost blood tests with artificial intelligence (AI).
The hope is that the combination can help catch cancers as early as possible when there is a greater chance of cure, and so far there’s been enough progress to nudge biotech firms to bet on the potential. In 2016, for example, DNA sequencing giant Illumina announced the formation of California-based GRAIL. With big-name backing like Bezos and Gates, the plan is to combine genomic screening and computational algorithms to assess clinical data and pinpoint cancer-specific patterns.
To get there, GRAIL is conducting several population-scale clinical studies. Its STRIVE Study, originally designed to develop and evaluate a breast cancer detection test, has enrolled roughly 100,000 women without a known cancer diagnosis. As Vice President of Clinical Development Anne-Renee Hartman explained in November, findings from the Circulating Cell-free Genome Atlas Study have led to a shift in focus as the company looks to advance a multi-cancer blood test.
Meanwhile, Freenome—another company based in The Golden State—is developing blood tests that look beyond tumor mutations to detect the body’s early-warning signs of cancer, including changes in immune system activity. The intent is to use AI to recognize disease-associated patterns among circulating, cell-free biomarkers. In October, the company revealed early data on efforts toward a colorectal cancer (CRC) screening test.
“We have shown that it’s possible to take a machine learning–based approach to decode the relationship between a patient’s cell-free DNA profile and his or her cancer status, detecting CRC in our dataset with a top performance of 82% sensitivity at 85% specificity,” Chief Medical Officer Girish Putcha said in a statement. “These are very encouraging data that support the continued development of a CRC screening test that includes cfDNA and machine learning as key components.”
SEEKing Improved Detection
Academia isn’t necessarily sitting on the sidelines. University of Pennsylvania bioengineer Dave Issadore and colleagues published a paper in Cancer Research this past year (doi: 10.1158/0008-5472.CAN-17-3703) that explores the possibility of detecting pancreatic ductal adenocarcinoma at earlier, curative stages.
The study, which uses a mouse model, reports on methods used for isolating extracellular vesicles (EV) from plasma and reveals a workflow for identifying EV miRNA biomarkers using RNA sequencing and machine-learning algorithms. The findings, according to the authors, provide strong proof-of-concept support for applying this approach to liquid biopsy.
Another research team led by Nickolas Papadopoulos, a professor of oncology and pathology at Johns Hopkins University School of Medicine, has created a blood test called CancerSEEK that can detect eight common cancer types and help identify a tumor’s location. Their study, published in Science in 2018 (doi: 10.1126/science.aar3247), used this test on about 1,000 patients known to have nonmetastatic cancers of the ovary, liver, stomach, pancreas, esophagus, colorectum, lung, or breast.
CancerSEEK detected cancer roughly 70% of the time. Although the test was most successful at finding stage II or III cancers, it managed to detect over 4 in 10 stage I cancers. The sensitivity ranged from a high of 98% for ovarian cancer to a low of 33% for breast cancer, and the specificity was greater than 99%. Of the 812 healthy controls, 7 scored a false positive. The method incorporates circulating tumor DNA and protein biomarkers, as well as a machine learning approach that helps narrow down the location of a tumor to one of two organs in over 8 in 10 patients with a positive blood test result.
The next step, Papadopoulos told Diagnostics World News, is finding out how well the test works in terms of detecting cancer in the general population among those with no known history of the disease. This work is already underway in cooperation with Geisinger Health System. The first portion of the study, focused on the test’s specificity, includes 10,000 individuals ages 65 to 75. If successful, about 40,000 individuals will then be followed for a period of three to five years.
While this research is in the works, there are also efforts to improve the performance of the test and reduce the cost in order to develop a simple test that could someday be incorporated into a routine physical exam. “The idea of early detection is to try to catch cancers as early as possible, before people have symptoms, because we know that most cancers get diagnosed at late stage and then usually you talk about survival of months and perhaps years,” he said, as opposed to talk of cure. “So that is the goal, really, of this research is to change that.”
Monitoring Treatment Response
There are also efforts underway that extend beyond cancer screening. A variety of labs and companies have instead turned their attention to tracking the evolution of tumors and the response to treatment.
Dan Landau’s lab at Weill Cornell Medicine, for example, revealed a machine-learning method geared toward detecting cancer mutations in very low-quantity cell-free DNA in order to monitor treatment. This fall, Landau received a New Innovator Award from NIH. With the $1.5 million prize money, he plans to investigate how tumor cells from chronic lymphocytic leukemia evolve to evade therapy using liquid biopsy technology.
Swiss-based analytics company SOPHiA GENETICS, which revealed its AI-powered solution for liquid biopsies at the 2017 Annual Meeting of the American Society of Clinical Oncology, utilizes the company’s AI solution, SOPHiA, to accelerate early detection of cancer and track treatments’ effectiveness. The application for liquid biopsies is also available for clinical trials, to help identify patients most likely to respond to new treatments.
Using liquid biopsies to help guide cancer therapy choice is the focus of Cambridge, UK-based Cambridge Cancer Genomics (CCG). Using simple blood draws, CCG looks to shorten the time required to know whether a treatment is working, which potentially allows physicians added opportunities to make strategic adjustments along the way.
The average time for a tumor to be picked up on a PET/MRI or CT scan is about seven months, Cofounder and CEO John Cassidy told Diagnostics World, and his company claims it can identify relapse seven months earlier than this standard practice by picking up a signal from the DNA in a patient’s blood. This information was found from a group of 120 lung cancer patients analyzed during therapy, he said, although the findings have not yet been published.
“Because we are interested in following tumors over time to see if the drugs are working or not, it turns out that we get this really rich dataset,” Cassidy said. “If you use this longitudinal data, you can actually start to predict, using machine learning, in which direction the tumor is evolving in response to therapy.”
Once it is known how a tumor will evolve, it becomes possible to begin to understand which therapies could be useful six months or a year down the line. From there, “you can start helping doctors tee up these clinical trials and tee up these second or third-line therapies for a patient who has got a very rapidly evolving tumor,” he added.
Going forward, CCG is starting to build up what Cassidy claims will be one of the world’s largest datasets centered on tumor evolution and response to therapy. Coupled with clinical trial data and information on which investigational drugs work in which circumstances, the ultimate goal is to tailor an entire treatment process for an individual patient.
Not There Yet
While liquid biopsy and AI—and their combinatorial powers—may be of great interest, some experts temper all the anticipation of what could be with the current reality, noting that all this potential hasn’t been realized just yet.
Advanced computational methods may be able to help us in the future, said Anthony Magliocco, a pathologist at Moffitt Cancer Center in Tampa, Florida whose research focuses on the molecular mechanisms of cancer progression and the development of drug resistance, but he hasn’t seen AI take on any significant role in pathology to date.
“Certainly there’s a lot of hope for it,” he told Diagnostics World News, “but it’s not really been fully implemented.” And liquid biopsy is only minimally used in clinical settings. It’s an area with great promise, he added, “but the applications are very, very few and far between at the moment.”
Paul Nicolaus is a freelance writer specializing in science, nature, and health. Learn more at www.nicolauswriting.com.