The clinical challenge
Liquid biopsies have undoubtedly reshaped precision oncology. A simple blood draw can now reveal the molecular blueprint of a patient’s cancer, and this information can be subsequently used to guide treatment decisions. Actionable mutations, resistance mechanisms, and tumor evolution can all be evaluated without the need for an invasive tissue biopsy. But as the field has matured, so has our appreciation for a persistent confounder lurking in the very same blood sample we draw: clonal hematopoiesis.
As we age, hematopoietic cells accumulate somatic mutations, some of which are in cancer-related genes and can lead to increased cell proliferation. This phenomenon is called clonal hematopoiesis (CH) and is remarkably common. By some estimates, over 40% of individuals with advanced cancers harbor at least one CH variant among clinically reportable genes, and the prevalence increases steeply with age.
The problem? Many of these CH-derived mutations sit in the very same genes we use to match patients to targeted therapies: BRCA1/2, ATM, TP53, and others. To put this in clinical context, when a variant is detected in a patient’s blood, the clinician must determine whether it is a bona fide tumor alteration that warrants targeted therapy or a confounder arising from a hematopoietic clone unrelated to the cancer itself. Misinterpreting this could result in administering the incorrect medication.
Ways to alleviate these challenges include matched tumor or white blood cell (WBC) sequencing. If the variant is present in both circulating free DNA (cfDNA) in plasma and WBC DNA, it provides strong evidence that the variant is CH-derived. However, in routine clinical practice, tumor or WBC sequencing adds cost, complexity, and turnaround time. Consequently, physicians may need to make treatment decisions based on detected cfDNA variants, without complete confidence that those variants are originating from the cancer they are trying to treat.
This issue is tackled using a new machine learning model named plasmaCHORD.
What is plasmaCHORD?
plasmaCHORD (plasma Clonal Hematopoiesis ORigin Detection) is a machine learning model designed to determine, for each variant detected by a cfDNA targeted fixed-gene panel sequencing, whether it originated from tumor cells or from WBCs. PlasmaCHORD combines information across three levels:
1. Mutant cell-free DNA fragment-level features: structural characteristics of the cfDNA fragments carrying the variant, leveraging the differences in physical properties between tumor-derived and hematopoietic-derived cfDNA.
2. Variant-level features: properties of the mutation itself, including its genomic context and known prevalence in CH versus solid tumors.
3. Gene-level features: evaluation of whether the gene carrying the mutation is frequently mutated in CH versus solid tumors.
3. Patient-level features: clinical characteristics that may influence the probability of CH, such as age of the patient.
By combining these inputs, plasmaCHORD captures the complexity of cfDNA biology in the context of a specific host in a way current rule-based filters cannot.
How we built and validated it
We developed plasmaCHORD using a training set of 426 variants identified by next-generation sequencing of cfDNA from 225 patients with stage I–IV solid tumors. Critically, the origin of each variant, tumor, or CH was determined by matched sequencing of WBCs and tumor tissue. After fine-tuning the model parameters, we tested plasmaCHORD on an independent validation cohort of 1,418 plasma variants from 114 patients with metastatic cancers. The model maintained strong performance, with an area under the curve (AUC) of 0.94 in the training set and a similar AUC of 0.90 in the validation set. Importantly, plasmaCHORD demonstrated a significant improvement in accuracy, particularly for clinically significant genes that matter most for treatment decisions.
From computation to bedside
But what might the clinical impact of plasmaCHORD look like? We applied plasmaCHORD to cfDNA sequencing data from patients enrolled in a prospective precision oncology clinical trial (NCT05585684). We highlighted two challenging cases in which actionable variants were detected but the origin was ambiguous, raising the possibility that targeted therapy could be inappropriate or ineffective. PlasmaCHORD correctly classified these variants as tumor- or CH-derived, helping to avoid inappropriate genotype-based treatment decisions and to accurately identify patients who might otherwise have been incorrectly matched to a targeted therapy.
The bigger picture
The challenge of CH in liquid biopsies is not going away; if anything, it is growing. As liquid biopsy panels expand to cover more genes, as comprehensive genomic profiling moves into earlier disease stages, and as the population of cancer patients ages, the risk of CH-driven misinterpretation will only increase. Approaches like plasmaCHORD can bridge that gap, adding a layer of intelligent interpretation to the thousands of plasma-only cfDNA sequencing assays conducted each year. PlasmaCHORD is therefore a computational safeguard, ensuring that when we act on a variant, we are acting on the right signal.
Read the full paper here.