Methylation-based classification of circulating tumor DNA using enzymatic methyl-seq in a pan-cancer clinical cohort

Circulating tumor DNA (ctDNA) offers a minimally invasive approach for cancer detection and monitoring. The addition of DNA methylation-based biomarkers into liquid biopsy workflows has the potential to enhance analytical performance. We evaluated the performance of the NEBNext ®  Enzymatic Methyl-seq (EM-seq) kit for ctDNA-based cancer classification in a pan-cancer clinical cohort. A cohort consisting of 99 plasma-derived cfDNA samples (29 healthy donors; 70 cancer patients across five organ sites) were processed using a semi-automated EM-seq workflow and sequenced to a target depth of 10X. Multiple machine modeling methods were evaluated using DNA methylation ratio values segmented into bin sizes ranging from 10 bp to 350 bp, increasing in 10 bp increments. For each window size, a leave-one-out cross-validation strategy was applied to differentiate cancer from healthy samples. This resolution study resulted in the testing of over 31,000 models, and the predictive capacity of models was evaluated using sensitivity, specificity, and area under the curve (AUC). For each model, performance was evaluated by assessing the ability of the model to differentiate cfDNA from healthy donors and cancer patients. Logistic regression with a bin size of 290bp achieved the highest performance with an AUC of 0.98, sensitivity of 0.96, and specificity of 0.9. Model agnostic feature selection identified 15 features that highly overlap amongst cross-validation sets. Some bins mapped to known oncogenes such as NOTCH2 and CASC15 , or regions known to contribute to cell growth and tumor progression, including TGFA (a proto-oncogene), regulatory RNA LINC00665, and genes implicated in cancer biology such as HS3ST5 , SPOCK3 , ATG13 , and ERGIC1 . Other bins fell in intergenic regions or within genes with no known roles in oncogenesis ( EPS15L1 , ARMC9 ). The misclassified samples (n = 6; 3 healthy and 3 cancer) had a lower average mapping efficiency (mean = 87.5%) than correctly classified samples (mean = 94.5%; p = 0.003), suggesting that sequencing quality may influence classification outcomes. Different combinations of modeling methods and resolution levels (bin size) were also highly predictive of disease status. Methylation-based classification of ctDNA using EM-seq enables high-sensitivity cancer detection across multiple tumor types. Ongoing improvements in conversion efficiency, consistency across workflows, and expansion of sample cohorts may further enhance model robustness. These findings support the potential of methylation-informed liquid biopsy as a diagnostic tool in oncology.