用于早期癌症检测的基于cfDNA的全基因组多特征分类器的分析精密度与跨平台可复现性

Introduction: Genome-wide liquid biopsies offer a non-invasive approach to cancer detection, management, and monitoring, and are increasingly being adopted across the cancer care continuum. These approaches interrogate a multitude of genomic features of the cell-free DNA (cfDNA) fragmentome, providing potential performance advantages over traditional imaging and other blood-based biomarkers. As cfDNA-based fragmentome methods move towards clinical deployment, validating their analytical precision and replicability across sequencing platforms is essential to realize their full clinical potential. Methods: We developed an automation-compatible two-step plate protocol that combines end repair and A-tailing to process 48 samples in parallel. To systematically assess the analytical precision of this protocol, we designed a precision study using large volume (~20 mL) plasma from 12 individuals without cancer and individuals with lung (n=10) or liver (n=2) cancer. Three unique pre-specified plate layouts ensured that there was a minimum of 12 genomic library replicates per donor, with distinct row and column assignments within and between plates. In parallel, cfDNA was processed using our previously published three-step tube protocol, with a minimum of 6 genomic library replicates across three unique tube layouts. Each operator processed two unique plate and two unique tube layouts, with 3 ng input cfDNA for all libraries. Results: Library yields were higher with the tube-based three-step protocol compared to the plate-based two-step protocol (6.51 nM vs 5.25 nM). Presequencing analyses revealed measurable operator effects in library QC metrics across matched replicates. All libraries were sequenced genome-wide to a mean depth of 1-2x on both the NovaSeq 6000 and NovaSeq X. Post-sequencing, locked computational pipelines for alignment and feature summarization were applied uniformly, generating genome-wide fragmentation in non-overlapping 5 Mb bins via short (100-150 bp) to long (151-220 bp) fragment ratios (S/L ratios), chromosome arm-level measures of aneuploidy, and lung and liver cancer classifier scores. The intra- and inter-batch precision of these features and classifier scores was assessed by implementing a Bayesian hierarchical model using the Stan software. Conclusions: This study aims to provide a detailed assessment of assay- and platform-driven variability in cfDNA-based features and classifiers. Initial findings support the feasibility of scalable, automation-compatible plate-based protocols, and emphasize the need for robust workflows in clinical implementation. Ongoing analyses will determine the stability of classifier performance across protocol and sequencing platform transitions, informing best practices for evolving liquid biopsy technologies.