AI Model Improves Liquid Biopsy Accuracy by Filtering Biological Noise

A new machine learning model, plasmaCHORD, enhances the precision of liquid biopsies by distinguishing between tumor-derived DNA and DNA originating from white blood cells. Developed by researchers at the Johns Hopkins Kimmel Cancer Center and published in Clinical Cancer Research, the tool addresses the common challenge of clonal hematopoiesis, where aging-related white blood cell mutations can mimic cancer signatures in blood samples.

How Clonal Hematopoiesis Affects Biopsy Results

Liquid biopsies are non-invasive tests that analyze cell-free DNA (cfDNA) circulating in the bloodstream to guide cancer treatment. However, clinicians often face “biological noise” when these tests detect mutations that do not originate from a tumor. According to Dr. Jenna Canzoniero, an assistant professor of oncology at the Johns Hopkins University School of Medicine, these mutations frequently arise from clonal hematopoiesis—a process where white blood cells acquire mutations due to aging, prior chemotherapy, or radiation exposure.

Without a way to verify the source, a clinician might mistakenly identify a white blood cell mutation as a tumor marker, potentially leading to the prescription of an ineffective targeted therapy. By filtering out these non-tumor-related variants, plasmaCHORD helps ensure that treatment plans are based on accurate genomic data.

The Mechanics of plasmaCHORD

The plasmaCHORD model utilizes the distinct physical characteristics of DNA fragments to differentiate their origins. Research indicates that DNA from tumor cells and white blood cells undergo different “chopping” processes, resulting in unique fragmentation profiles. The model analyzes these patterns alongside clinical factors, including the patient’s age and specific gene mutation types, to determine the likelihood that a mutation is tumor-derived.

In a validation study, the team trained the algorithm using 225 patient samples across various cancer types, including breast, colorectal, and non-small cell lung cancer. By comparing these results against matched genetic sequencing of both tumor and white blood cells, the researchers confirmed the model’s ability to isolate genuine cancer mutations.

Clinical Impact and Performance

Testing the model on an independent cohort of 114 patients demonstrated a significant improvement in diagnostic accuracy. According to the study findings, plasmaCHORD increased the success rate of distinguishing tumor mutations from white blood cell mutations from 50% to 83% for clinically relevant variants. This refinement is critical for the Johns Hopkins Molecular Tumor Board, where the tool has already provided proof-of-concept evidence in helping clinicians avoid therapies that would likely be ineffective.

Dr. Valsamo Anagnostou, senior study author and leader of the Johns Hopkins Molecular Tumor Board, notes that approximately one-third of mutations detected in liquid biopsies can originate from white blood cells. Because the model is designed to work with standard liquid biopsy data, it offers a scalable solution for integrating more precise genomic profiling into routine clinical oncology.

Frequently Asked Questions

• What is a liquid biopsy? It is a blood test that detects fragments of DNA shed by tumors, allowing doctors to identify genetic mutations without a surgical tissue biopsy.
• Why do white blood cell mutations complicate test results? As people age, white blood cells can develop mutations that look similar to cancer mutations on a genomic report, which can mislead clinicians regarding the tumor’s actual genetic makeup.
• Is plasmaCHORD available for clinical use? The model is currently a research tool, though its developers suggest it is designed to be quickly scalable for clinical settings to help verify the origin of mutations in future testing.

Key Takeaways

• Problem: Up to one-third of mutations found in liquid biopsies may be “biological noise” from white blood cells rather than the patient’s tumor.
• Solution: The plasmaCHORD machine learning model uses DNA fragmentation profiles to identify the true origin of these mutations.
• Result: Clinical accuracy in identifying relevant tumor mutations improved from 50% to 83% in independent testing cohorts.