Abstract 3366: A machine learning approach to determine the cellular origin of variants in liquid biopsies

Introduction: Targeted next-generation sequencing (NGS) of cell-free DNA in plasma, referred to as liquid biopsy, has become a valuable diagnostic tool in clinical oncology. However, detection of variants related to clonal hematopoiesis (CH) is a major confounder that significantly impairs the clinical utility of liquid biopsies. Here we developed a machine-learning model to determine tumor versus CH origin of variants identified in plasma-only NGS.

Methods: We assembled a training cohort of 352 variants identified by targeted deep plasma sequencing from 199 patients with stage I-IV breast, colorectal, esophageal, lung, and ovarian cancer, coupled with matched white blood cell (WBC) and tumor tissue NGS to allow determination of the reference origin for each plasma variant. We employed Extreme Gradient Boosting (XGBoost) to integrate fragment, variant, gene, and patient level features to predict tumor versus CH plasma variant origin, evaluating the performance of this approach within the training cohort using 10-fold cross-validation. We applied the fixed model to two independent validation cohorts: a small cell lung cancer (SCLC) cohort comprising of 74 variants from targeted plasma NGS from 26 patients and a multi-cancer cohort of 409 variants detected using the MSK-Impact panel from 74 patients with breast, colorectal, and prostate cancer.

Results: Variant allele frequencies (VAF) did not differentiate tumor from CH variants, as the VAFs between tumor (median VAF 0.53%) and CH (median VAF 0.409%) variants in the training cohort were largely overlapping (area under the ROC curve-AUC 0.54, 95% confidence interval-CI 0.48-0.61). Similarly, individual fragmentomic features (mutant fragment length, cut points, and endpoint motifs) had limited ability to distinguish tumor from CH variants (AUC range 0.51-0.76). Using serial plasma samples, we identified stable statistical measures of differences in fragment feature distributions between mutant and wild type fragments; these were subsequently incorporated into an XGBoost machine-learning model along with variant, gene and patient features to predict tumor versus CH variant origin. Our model predicted variant origin with an AUC of 0.95 (95% CI 0.87-1) from 10-fold cross validation in the training cohort. The performance of the model was tested in independent SCLC and multi-cancer validation cohorts; the fixed model predicted plasma variant origin with an AUC of 0.87 (95% CI 0.73-1) and 0.89 (95% CI 0.86-0.92) respectively.

Conclusion: We developed a machine-learning model that integrates patient, gene, variant and fragment features to predict tumor versus CH origin of plasma variants across solid tumors and NGS sequencing platforms. The ability to identify bona fide tumor variants in plasma-only sequencing fills a critical need in the clinical implementation of liquid biopsy-guided cancer therapy by reducing misinterpretation due to CH contamination.

Citation Format: Jenna V. Canzoniero, Archana Balan, Jillian Phallen, Blair V. Landon, Lavanya Sivapalan, Benjamin Green, Zineb Belcaid, Susan C. Scott, Gavin Pereira, Vincent K. Lam, Ali H. Zaidi, Ronan J. Kelly, Christine L. Hann, Wade T. Iams, Christine M. Lovly, Patrick M. Forde, Gerrit A. Meijer, Geraldine R. Vink, Remond J. Fijneman, The MEDOCC Group, Victor E. Velculescu, Robert B. Scharpf, Valsamo Anagnostou. A machine learning approach to determine the cellular origin of variants in liquid biopsies [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 3366.