Ilias Ziakas

Supplementary Table S1: Training cohort description with plasmaCHORD feature values. Supplementary Table S2: Description of the serial cohort. Supplementary Table S3: Validation cohort description with plasamCHORD feature values. Supplementary Table S4: Description of plasmaCHORD features. Supplementary Table S5: Correlation of cfDNA fragmentomic feature statistics between baseline and follow-up time points in the serial cohort. Supplementary Table S6: Possible CH-origin variants identified on retrospective analysis of 45 patients reviewed at the JH-MTB.

Purpose:Targeted next-generation sequencing (NGS) of cell-free DNA (cfDNA) enables comprehensive molecular profiling and can guide the selection of genotype-targeted therapies. However, the detection of variants derived from clonal hematopoiesis (CH) is a significant confounder in liquid biopsies.Experimental Design:Using a training cohort of 426 variants identified in cfDNA NGS from 225 patients with stage I to IV solid tumors, we developed plasma Clonal Hematopoiesis ORigin Detection (plasmaCHORD), a machine learning model that includes fragment-, variant-, and patient-level features to distinguish between tumor and CH origin for each variant detected by liquid biopsies. Model performance was assessed by comparison with the reference origin for each plasma variant determined from matched white blood cell and tumor NGS. Following the locking of the model parameters, we applied plasmaCHORD to an independent validation cohort of 1,418 plasma variants detected in 114 patients with metastatic cancers, as well as to cfDNA NGS from patients enrolled in a prospective clinical trial (NCT05585684).Results:plasmaCHORD predicted tumor origin versus CH origin in the training set with high accuracy (AUC = 0.94). In the independent validation cohort, the locked model maintained similar overall accuracy (AUC = 0.9) and demonstrated significant improvement in accuracy for clinically significant genes. When applied to clinically challenging cases in the context of a precision oncology clinical trial, plasmaCHORD precisely determined variant origin, preventing mismatches with genotype-targeted therapies.Conclusions:plasmaCHORD, a multifeature machine learning model, can significantly enhance the ability to identify bona fide tumor variants in routine plasma-only NGS, addressing a critical need for implementing liquid biopsy–guided therapy by minimizing misinterpretation caused by CH.

Classification schema for determining reference origin of plasma variants using matched WBC and/or tumor tissue sequencing. For each variant detected in plasma, reference origin was determined by comparison with matched WBC sequencing and tumor sequencing (if available). If patient had both WBC and tumor sequencing, then plasma variants that were detected with at least 3 read families in WBC and not detected in tumor were considered WBC-derived. Variants were considered tumor-derived if they were detected with VAF >10% in tumor, or if the variant was detected by IGV in the tumor and was detected in 1 or fewer read families in WBC sequencing. If patient had only WBC sequencing, then plasma variants that were also identified in WBC sequencing with at least 3 read families were called CH-derived. Plasma variants that were not identified in matched WBC sequencing (0 read families) where there was adequate coverage in the WBC sequencing at that locus were called tumor-derived. Variants that did not meet any of these criteria were determined to be unknown origin and were excluded from the remaining analysis.

Performance of individual features in predicting variant origin in the training cohort. Receiver-operator curves and area-under the curve (AUC) values for individual model features as predictors for CH versus tumor-origin. The top performing individual features are gene heme fraction, length cluster mutant, and length max delta S. However, no individual feature performs as well as the plasmaCHORD model combining all features.

Assessment of fragmentomic features. (A-B) Illustrative examples of fragmentome distribution differences seen between mutant and wild-type reads from CH-origin versus tumor origin variants. (A) Tumor-derived plasma variant TP53 p.S215R with fragment length density left-shifted (shorter) for mutant compared to wild-type fragments (middle), and with the fragment endpoints relative to mutation location closer together and left-shifted in mutant versus wild-type fragments (bottom). (B) CH-origin variant TP53 p.M237I with overlapping fragment length densities (middle) and fragment endpoint locations (bottom) between mutant and wild-type reads. (C) The serial cohort used to assess fragmentomic features over time. Each column represents a variant, with points on the y-axis depicting VAF at baseline (black) and at follow-up (gray) time points. This cohort included patients with colorectal, esophageal, and NSCLC, stage I-III and included both tumor-origin and WBC-origin plasma variants. (D) Pearson’s product moment correlation of fragmentomic feature values at baseline and follow-up timepoints. Correlation of features between baseline and follow-up timepoint was calculated regardless of variant origin (black bars). Features that were correlated over time with p-value <0.05 were subsequently used for development of the machine learning model. Correlation of features between baseline and follow-up timepoint was also calculated for the subset of tumor-derived variants (red bars) and CH-derived variants (blue bars), demonstrating similar correlation of features over time regardless of variant origin.

Pearson correlation for each individual fragmentomic statistic between baseline and follow-up time points as determined using the serial cohort. Definition of summary statistics are provided in Supplementary Table S2. Each statistic is calculated by comparing mutant versus wild-type reads at a given locus. Statistics that were highly correlated between baseline and follow-up time points (regardless of mutation origin) were included as features in the machine learning model.

Description of the independent validation cohort. (A) Distribution of plasma variant allele frequency for tumor-origin and WBC-origin variants in the validation cohort. (B) Distribution of variant origin in the validation cohort was determined in the original publication. IMPACT-BAM-matched as defined in the original publication refers to variants detected in tissue biopsy below the limit for variant calling as required for clinical reporting. Note we combined biopsy-matched and IMPACT-BAM-matched as tumor-derived. In this cohort, reference origin for plasma variants as determined by matched sequencing is 63.8% CH-derived and 36.1% tumor-derived. (C) Distribution of alterations and variant origins in genes with more than 10 variants.

Model performance in the training cohort by number of mutant reads supporting the mutant DNA molecules. Receiver operator curve and AUC values for subsets of training cohort as determined by the number of mutant reads. As the number of mutant reads decreases, the margin of error in the estimated value of the summary statistics comparing distributions between wild-type and mutant reads increases. To evaluate the impact of this on the overall model performance, we determined the model performance in subsets of variants based on number of mutant reads: 3-5 reads inclusive, 6-10 reads, 11-20 reads, 21-100 reads, and >100 reads. For variants with only 3-5 mutant reads, the model performed slightly less well, with AUC = 0.841. For all subsets with >5 mutant reads, the model performed similarly well with AUC 0.89-0.939.

Description of the training cohort. (A) Distribution of cancer types, stage, and patient age in the training cohort. The mean age for patients in this cohort was 65 years old (range 19-87), the most common cancer type was colorectal, and most patients had early stage disease (22 stage I, 101 stage II, 48 stage III, 53 stage IV, and 1 unknown stage). (B) Distribution of WBC only versus WBC and tumor matched sequencing and determined reference variant origin. The proportion of variants with tumor-derived reference origin is slightly higher from samples that had both WBC and tumor sequencing versus WBC sequencing alone. (C) Distribution of variants included in training cohort by gene and reference variant origin. After DNMT3A, TP53 was the second-most commonly mutated gene (98 variants) of which 67 (68%) are tumor-derived and 31 are WBC-derived.