跨两种全基因组测序平台评估健康血浆样本中的测序错误率

Accurate detection of error signals in healthy samples is critical for improving assay specificity and sensitivity in ctDNA analysis. Sequencing artifacts and biological noise, arising from PCR, oxidative damage, or clonal hematopoiesis variants, can obscure low-frequency tumor somatic mutations. This study aims to systematically characterize error profiles in healthy plasma samples and develop context-aware filtering and models to distinguish true variants from background noise. Whole-genome sequencing data were generated from 80 healthy plasma samples (pWGS) using two sequencing platforms (sequencer A: N=52; sequencer B: N=28). Mutation-specific features were computed across sequencing fragments using a custom pWGS analysis module that incorporated information both at the fragment- and the position-level. We compared a probabilistic classifier (Model 1) and an advanced deep learning based model (Model 2) trained on the same feature set, incorporating features that we identified as key predictors of sequencing error. Error rates within the plasma samples were quantified at genomic positions corresponding to tumor mutation target sets identified from 26 tissue samples (breast N=4, lung N=5, ovarian N=10, N=7 bladder). Using sequencing platform A, error rates showed strong dependence on fragment-level and sequence-context features. Error rates were not evenly distributed among variant types. Error rates were elevated near fragment ends and within GC-rich regions (>55% GC). Median raw error rates across cancer indications ranged from 19 to 59 parts per million (ppm). Application of Model 1 reduced background error rates by up to 65%, while Model 2 achieved up to a 90% reduction, with minimum error rates approaching ~5 ppm. Sequencing data from platform B that were processed with a base quality > 50 filter achieved comparable coverage (~65×) and similar error rates (~5 ppm). Across both sequencing platforms, further reductions are possible by incorporating additional fragment-level features and leveraging more advanced modeling approaches. This work establishes a foundational framework for pWGS error characterization and demonstrates the effectiveness of fragment- and context-based modeling in reducing sequencing noise.