一种利用甲基化测序数据估计肿瘤分数的新方法

Introduction: Accurate estimation of tumor fraction in cell-free DNA (cfDNA) is critical for molecular residual disease monitoring and early cancer detection. Current approaches, such as maximum somatic allele frequency, i.e., MSAF, rely on the presence of detectable somatic mutations, limiting their sensitivity in samples with low tumor burden or shallow sequencing depth. Moreover, these approaches are susceptible to bias from clonal hematopoiesis and uneven mutation distribution, which can compromise the accuracy of tumor fraction (TF) estimation. DNA methylation signatures offer a promising avenue due to their abundance, consistency, and robustness. However, current methods to harness this information for tumor fraction estimation remain suboptimal. Methods: We developed a new and tumor-naive computational framework that leverages DNA methylation signatures to quantify tumor-derived DNA with markedly high analytical sensitivity. The approach integrates pre-selected methylation markers and weighted linear regression to quantify tumor-derived DNA to a minimal detection threshold. Our machine learning-based method combines tumor- and normal-specific prior methylation frequencies and target coverage information to achieve a more reproducible and reliable estimation of tumor fraction. The method was developed using in-house clinical samples, and its performance was benchmarked against mutation-based ctDNA tumor fraction from FoundationOne Liquid CDx (F1LCDx). Besides, in silico and in vitro dilutions were used to assess accuracy and sensitivity at low tumor fractions. Results: Our method demonstrated a limit of detection reaching the 0.01% TF, with strong accuracy maintained down to 0.1% in both in vitro and in silico dilution series. In vitro dilution experiments demonstrated a Spearman correlation of 0.86 and a Pearson correlation of 0.88 on log₁₀-transformed data, with median fold changes of 0.84 at 0.1% TF and 3.76 at 0.01% TF. Performance was further supported by in silico dilution analyses, yielding a Spearman correlation of 0.98 and Pearson correlation of 0.97 on log₁₀-transformed data. Additionally, when benchmarked against ctDNA TF estimates from F1LCDx, our method achieved a Spearman correlation of 0.88, a Pearson correlation of 0.90, and a median fold change of 0.91. Conclusions: We present a novel methylation-based approach for TF estimation in cfDNA. This approach enables highly sensitive and accurate TF estimation, particularly at low TFs, highlighting its potential utility in molecular residual disease detection, early cancer detection or treatment response monitoring. This approach offers a powerful tool for advancing liquid biopsy applications in cancer diagnostics and clinical management, with validation in larger cohorts ongoing.