Fig. 1: Study design and patient enrollment.

a An approach called ‘expanded multimodal analysis’ (EMMA) has been developed using machine learning to enhance the detection of ctDNA from cfDNA in plasma samples. This is achieved by comprehensively analyzing cancer-derived differentially methylated regions (DMRs), copy number variants (CNVs), and fragmentation features in the cfDNA whole-genome bisulfite sequencing (cfWGBS) data. The cancer-derived DMRs and CNVs were initially identified from paired WGBS and whole-genome sequencing (WGS) data of primary tumors and matched adjacent non-neoplastic tissues of 155 patients with esophageal squamous cell carcinoma (ESCC). Subsequently, the ESCC-derived DMRs and CNVs were examined in cfWGBS data and further utilized with the proportion of short cfDNA fragment sizes to train the diagnostic models in the discovery cohort. The performance of each diagnostic model was independently assessed in an external ESCC cohort and a precancerous cohort. To unveil the biological significance of these optimal DMRs, we correlated them with multi-omics-based molecular subtypes and transcriptomic profiles in the paired ESCC tissue samples. b The discovery cohort encompassed 150 patients with ESCC or high-grade intraepithelial neoplasia and 150 matched health controls to construct the diagnostic model using different cfDNA features. The performance of each diagnostic model was evaluated independently in an external ESCC cohort and a precancerous cohort. ESCC esophageal squamous cell carcinoma, IEN intraepithelial neoplasia, WGS whole-genome sequencing, WGBS whole-genome bisulfite sequencing, cfWGBS cfDNA WGBS, RNAseq RNA sequencing, HC healthy control, CNV copy number variant, DMR differentially methylated region, IM immune modulation, CCA cell cycle pathway activation, IS immune suppression, NRFA NRF2 oncogenic activation.