CancerScreen: A novel ultrasensitive liquid biopsy for early-stage cancer detection by ctDNA Duplex Sequencing and Tissue of Origin identification with supervised machine learning.

Abstract

60 Background: Circulating tumor DNA (ctDNA) in blood holds promise as a cancer-specific biomarker for early-stage cancer diagnosis. However, detection of ultra-low mutation allelic frequency (MAF) of ctDNA at early stages of cancer is infeasible by conventional next generation sequencing (NGS). Using duplex sequencing with unique molecular identifiers (UMIs) and custom-designed probes, we tested the hypothesis that ctDNA duplex sequencing with UMIs was able to detect ultra-low MAF of ctDNA in patients with early-stage cancers. Methods: A 128-gene panel that contains probes targeted to clinical relevant genome variations in cancers of the lung, stomach, and esophagus was designed and validated with reference DNA and controls using ctDNA duplex sequencing with UMIs. A data analysis pipeline was implemented withimproved algorithms for variant calling, blood tumor mutational burden (bTMB) calculation, and supervised machine learning for tissue-of-origin primary cancer identification. Results: We designed and validated a ctDNA duplex sequencing with UMIs assay that enables simultaneous detection of 128 clinical relevant geneswith SNPs, indels, amplifications, and fusions in a single blood test. Compared to conventional ctDNA NGS, our assay achieved high sensitivity (over 82%) and specificity (over 96%) with LOD at 0.1% MAF for stage I lung, gastric and esophageal cancers with the sequencing depth at 30,000x from a cohort of 136 clinical samples. Results also showed significant concordance of MAF and TMB between DNA from tumor tissues and plasma ctDNA. Our deep learning predictive model with novel algorithms and features for tumor tissue-of-origin classification achieved an overall 85% accuracy. Conclusions: In this study, a novel ultrasensitive assay was designed and validated for accurate detection of MAF at 0.1% from plasma ctDNA of multiple tumors, and accurate classification on tissue-of-origin for major primary cancers using supervised deep learning. The results of this liquid biopsy study from initial clinical testing showed its promise on clinical applications for early-stage cancer diagnosis.