When oncologists order a liquid biopsy to find the mutations driving a patient's cancer, they often face a frustrating puzzle: the test reveals genetic changes, but not their source. Johns Hopkins researchers have now built an artificial intelligence model that solves this mystery with remarkable precision, helping clinicians prescribe the right drug to fight the right mutations.
Liquid biopsies—blood tests that capture fragments of cell-free DNA from tumors—have transformed cancer care by revealing the specific mutations that make each patient's tumor unique. Doctors use these insights to match patients with targeted therapies designed to attack those exact genetic weaknesses. But there is a catch. The blood samples also pick up mutations from white blood cells, which accumulate over time through an aging-related process called clonal hematopoiesis. This biological noise is especially common in older patients and those who have previously received chemotherapy or radiation. When a clinician sees a mutation on a liquid biopsy report, they cannot tell whether it came from the tumor or from background white blood cell mutations—a critical distinction when selecting which drug to prescribe.
To cut through this ambiguity, a team led by Jenna Canzoniero, M.D., M.S., at the Johns Hopkins Kimmel Cancer Center developed plasmaCHORD, a machine learning model that distinguishes tumor mutations from white blood cell mutations by analyzing the distinctive patterns in how different DNA fragments are "chopped up" in the blood. Tumor DNA and white blood cell DNA create distinct fragmentation profiles, Canzoniero explains, and the model learns to recognize them alongside other clinical clues like the patient's age and the type of mutation involved.
The researchers trained plasmaCHORD on liquid biopsy samples from 225 patients with breast, colorectal, esophageal, ovarian, or non-small cell lung cancer, verifying the model's accuracy against matched genetic sequencing of patients' actual tumor cells and white blood cells. They then tested it on a completely separate cohort of 114 patients with breast, prostate, or non-small cell lung cancer from another institution using a different sequencing platform—a stringent validation that confirmed the model's real-world reliability. For clinically relevant mutations, plasmaCHORD improved accuracy from approximately 50 percent to 83 percent.
Perhaps most tellingly, when Johns Hopkins clinicians presented plasmaCHORD's predictions to their Molecular Tumor Board, the model helped them avoid recommending therapies that would likely have failed. About one-third of all mutations detected in liquid biopsies originate from white blood cells rather than tumors, notes Valsamo Anagnostou, M.D., Ph.D., the senior author and leader of the Johns Hopkins Molecular Tumor Board. The research, published May 1 in Clinical Cancer Research, demonstrates that a standard AI tool applied to routine liquid biopsies could be both clinically valuable and easily scalable across institutions.
Canzoniero says the team is already planning next-generation versions of plasmaCHORD aimed at even higher accuracy, but the current model is ready for use in research and clinical settings today. For patients with cancer, that means one less layer of uncertainty when clinicians choose their next treatment.