A multiomic ensemble-based approach for high-specificity detection of clonal hematopoiesis in cfDNA liquid biopsy

Introduction: Variants arising from clonal hematopoiesis (CH, or CHIP) can be misclassified as tumor-derived in liquid biopsy, and are a challenge especially in expanded gene panels. Sequencing both buffy coat and plasma samples can help adjudicate CH but is costly and limited by the sensitivity of buffy coat sequencing at low allele fractions and its inability to capture somatic variants of non-tumor and non-peripheral blood origin. To address this we developed a high-specificity plasma-only CH classifier for late-stage cancer patients and applied it to clinical data for insights into CH prevalence and potential clinical impact. Methods: This multiomic classifier categorizes non-germline variants as CH- or tumor-derived and is an ensemble model leveraging fragmentomics, methylation tumor fraction (mTF), variant allele frequency (VAF) and variant- and sample-level metadata from over 250,000 Guardant Health samples and public data bases (GnomAD, COSMIC). A total of 1033 samples consisting of paired buffy coat and plasma (N=686) or paired tissue-plasma (N=305) from patients with cancer, and samples from cancer-free individuals (N=42) were used for training and testing, reflecting the age (median: 65) and cancer type distribution of the intended use population. Population-level CH characteristics were evaluated in 30,000 late-stage cancer samples (median age: 67). Results: Feasibility results of 185 testing samples showed panel-wide sensitivity and specificity over 90% in genes with recurrent CH variants (N≥6). In nearly 30,000 late-stage cancer samples, CH variant rates were elevated in prostate cancer relative to other cancer types. Somatic variant counts increased with rising mTF, whereas CH variant counts remained relatively stable. In samples with mTF of 1% or higher, our approach revealed that for 16.9% of samples the maximum variant allele frequency (VAF) corresponded to a CH variant. Finally, we recapitulated previously reported findings that lymphoid-associated CH variants increase less rapidly with age compared to myeloid-associated CH variants. Conclusion: We present a plasma-only CH classifier, applied on cancer patients run on the Guardant360 Liquid test (GuardantHealth, Palo Alto, CA). Feasibility data shows panel-wide sensitivity and specificity above 90%, enabling population-scale characterization of CH. Notably, in almost 17% of samples with mTF of 1% or greater the lead variant was determined to be non-tumor-derived, highlighting that this approach allows for disambiguation of CH variants that contribute to total VAF in cfDNA (maxMAF).