Modern cancer medicine is hampered by two big challenges—detecting cancers when they are small and offering cancer patients personalized, dynamic cancer care. To find solutions, several academic labs and biotech firms are turning to artificial intelligence, working to develop machine-learning algorithms that could help decipher weak signals in the blood that can identify cancers at an early stage and indicate whether a cancer is responding to treatment in real time.
“You have to find this needle in a haystack . . . this very weak signal amongst all of the cacophony of everything else happening in the body,” says Dave Issadore, a bioengineer at the University of Pennsylvania and founder of Chip Diagnostics, which is developing a machine-learning method to diagnose disease by sequencing extracellular vesicles’ cargo.
So far, machine-learning algorithms designed to detect minute quantities of tumor DNA in a blood sample—the goal of so-called liquid biopsies—have performed well in clinical validation studies, but no self-learning algorithm has yet been approved for clinical use. These have the potential to outperform imaging and tissue biopsies in detecting and monitoring cancers by looking for mutations in DNA, RNA, and proteins directly from the blood.
“I think the dream of liquid biopsy is to, A), detect cancer when there is very little of it, and, B) detect cancers during treatment,” says Dan Landau, a clinical oncologist and researcher at Weill Cornell Medical School in New York. However, in the both contexts, liquid biopsy techniques struggle to accurately detect cancer among the infinitesimally small quantities of tumor nucleic acids in the blood. Although the technique’s performance varies between cancer types, liquid biopsies so far have been able to detect cancer in around half of early-stage patients diagnosed through imaging, giving it a sensitivity of just 50 percent.
Machine learning algorithms need more time to refine and learn from more clinical evidence from the trials.—Siew-Kee Low, Japanese Foundation for Cancer Research
The problem is that “we do not have the luxury to run all the different kinds of assays with the limited cell-free nucleic acid,” writes Siew-Kee Low, a researcher at the Japanese Foundation for Cancer Research in Tokyo, in an email. Because of this limitation, only targeted assays of known cancer mutations have so far been possible. But many cancers lack a simple set of known mutations that can identify a tumor and, subsequently, determine the best treatment option. That’s where machine-learning comes in.
Artificial neural networks, which use thousands of connected nodes to interpret data much like neurons in the brain and form the basis of machine learning, can process vast amounts of data and identify patterns that would likely elude a human doctor. Not only that, machine learning is self-improving; as more data are put into the system, it fine-tunes its own algorithm to improve its diagnostic acumen. “It’s an evolving diagnostic,” says Gabe Otte, CEO of artificial intelligence (AI) genomics company Freenome. The company announced its first clinical validation study in May on a machine-learning algorithm that screens at-risk patients for DNA, RNA, and protein sequences in the blood that indicate colorectal cancer.
Machine learning is only ever as good as the data it has been trained on—feeding algorithms extensive, representative samples of DNA, RNA, or other biomarkers from which to learn is crucial to creating a highly sensitive test. To address this issue, cancer genomics company GRAIL has undertaken a large-scale study, the Circulating Cell-Free Genome Atlas (CCGA), that uses machine learning to create a huge, representative library of cancer mutations and healthy mutations using data from cfDNA and white blood cell genomes, on which to train their cancer screening algorithms. The company published initial results at the 2018 American Society of Clinical Oncology (ASCO) Annual Meeting this month. “The CCGA study is designed to be representative of a real-world population,” says Anne-Renee Hartman, vice president of clinical development at GRAIL. The study has enrolled 70 percent of the planned 15,000 participants, which will include those with one of 20 types of cancer and at all stages as well as cancer-free controls and participants suffering from other common diseases such as diabetes.
“Machine learning algorithms need more time to refine and learn from more clinical evidence from the trials and multi-layered datasets in order to provide good prediction for clinical decision making,” writes Low.
You can’t PET-MRI a tumor every week, but you could take a liquid biopsy.—John Cassidy, Cambridge Cancer Genomics
With improved power to detect cancer at low levels, “oncology will become a much more iterative field, where doctors are guided in how to treat a tumor in real time,” says John Cassidy, cofounder and CEO of Cambridge Cancer Genomics, a startup developing artificial intelligence to analyze mutations in cfDNA from multiple liquid biopsy tests to evaluate the effectiveness of a patient’s current treatment as it progresses. “Each tumor is unique, but the processes which drive this heterogeneity don’t stop during treatment: tumors are dynamic and ever-evolving,” he says. The team says its algorithm can detect cancer relapse up to seven months earlier than imaging when tested on publicly available datasets (the data are not yet published). The company announced the first phase of clinical testing of the software at The European Oncology Convention this month.
Earlier in the year, Landau’s lab announced a new machine-learning method to detect cancer mutations in very low-quantity cell-free (cf) DNA, with the hope of applying it to monitoring cancer treatment. The algorithm compares whole-genome sequences from tumor biopsy samples with patterns of mutations in fragments of cfDNA extracted from the blood. Scientists normally identify real tumor mutations from sequencing errors by measuring how many of the millions of repeated fragments of DNA contain a given mutation—the more fragments agree, the more certain you can be that the mutation is present in the tissue. But with so few fragments of cfDNA to work with, Landau’s software instead looks for complex patterns of mutations across the entire sequence to estimate whether a fragment has been sequenced correctly. Using this method, the algorithm detected non-small cell lung cancer mutations with 90 percent sensitivity in two patients, considerably better than standard liquid biopsy techniques perform.
Techniques combining liquid biopsy with machine learning will not only be less invasive than traditional diagnostic methods such as imaging and tissue biopsy, they may also be less expensive, sources tell The Scientist. “AI genomics cancer screening methods will be cheaper than screening procedures like mammography or colonoscopy,” says Otte. And AI monitoring of liquid biopsy data throughout treatment could also be cheaper, says Cassidy. “You can’t PET-MRI a tumor every week, but you could take a liquid biopsy.”
Correction (June 28): Dave Issadore’s affiliation is the University of Pennsylvania, not Penn State. The Scientist regrets the error.