A machine learning model developed by researchers at the Johns Hopkins Kimmel Cancer Center filters out biological noise in liquid biopsy samples, helping clinicians better tailor treatments to a patient’s tumor. These findings were presented in the following paper by Canzoniero et al. clinical cancer research.
Liquid biopsies, which analyze tumor cell-free DNA (cfDNA) fragments in blood samples, are commonly used to identify mutations in solid tumors in patients, allowing clinicians to select mutation-targeted therapies. However, liquid biopsies can also detect mutations that accumulate in white blood cells through an age-related process called clonal hematopoiesis. These white blood cell mutations are more common in older people and in patients who have previously received chemotherapy or radiation therapy.
“Even if we do a liquid biopsy and the report comes back and they find a mutation, we don’t know whether the mutation is coming from the tumor or from the white blood cells,” he explained. Jenna Canzoniello, MD, MSco-first author of the paper and assistant professor of oncology at Johns Hopkins University School of Medicine. “If you want to choose a drug that targets mutations to treat cancer, you need to make sure you’re targeting mutations in the cancer and not mutations in white blood cells.”
plasma cord
To solve this problem, Dr. Canzoniello and colleagues in the Molecular Oncology Laboratory developed a machine learning model called plasmaCHORD that uses features of DNA fragments to infer whether mutations found in liquid biopsies originate from tumors or white blood cells. Tumor DNA fragments and white blood cell DNA fragments are “chopped up” in different ways to create distinct “cfDNA fragmentation profiles,” Dr. Canzoniello explained. The model also uses factors such as the patient’s age and the type of gene or mutation.
The researchers trained their model using liquid biopsy samples from 225 patients with breast, colorectal, esophageal, ovarian, or non-small cell lung cancer. They validated the model’s accuracy by using matched genetic sequences from patients’ tumor cells and white blood cells to determine the true cause of the mutations. We then tested plasmaCHORD on another set of 114 patients with breast, prostate, or non-small cell lung cancer from another institution using a different type of liquid biopsy sequencing platform and found that the model had a similar ability to identify the true cause of mutations. Notably, within that cohort, plasmaCHORD improved its ability to accurately distinguish between tumor and leukocyte mutations from approximately 50% to 83% for a set of clinically relevant mutations.
Finally, they showed that predicting the origin of mutations with plasmaCHORD can help clinicians avoid selecting potentially ineffective treatments for patients evaluated by the Johns Hopkins Molecular Tumor Board, providing proof of concept that the information is clinically useful.
“About one-third of the mutations detected in tumor naive liquid biopsies can originate from white blood cells, and our ability to match targeted therapies to each patient’s genomic profile depends on our ability to distinguish tumor mutations from biological noise,” said Valsamo Anagnostou, MD, senior study author and Alex Glass Professor of Oncology and leader of the Johns Hopkins Molecular Tumor Committee at the Johns Hopkins University School of Medicine. “Applying artificial intelligence models to standard liquid biopsy tests has the potential to be clinically valuable and quickly scalable.”
“PlasmaCHORD can now be used for both research and potentially clinical purposes to determine the cause of mutations with liquid biopsy when in doubt,” said Dr. Canzoniello. “We are thinking of working on future versions that will hopefully have even better performance.”
Disclosure: For full study author disclosures, please visit aacrjournals.org/clincancerres.
