Fig. 2

The workflow of data analysis and models development. a The images of standard mammography and ultrasonography were interpreted and classified according to the fifth edition of the Breast Imaging Reporting and Database System (BI-RADS) standard by two experienced radiologists independently at each center. The cfDNA methylome of each participant was measured by the whole-genome bisulfite sequencing (WGBS). Additionally, machine learning was applied to identify cfDNA methylation markers of early-stage breast cancers. Finally, a diagnostic model using the identified cfDNA methylation markers in combination with radiology and ultrasound findings was accessed the ability to improve the accuracy of early-stage breast cancer detection. DCIS, ductal carcinoma in situ. b A comprehensive framework was devised to develop the classifier for identifying early-stage breast cancer based on the cfDNA methylation markers from massive cfDNA WGBS fragments. It includes four processes: the differentially methylated region (DMR) calling, cfDNA enrichment, cfDNA origin inference, and model development. First, 5613 hyper-DMRs and 51,962 hypo-DMRs were identified from WGBS data of 10 benign and 9 malignant breast primary tissue samples. Second, the cfDNA enrichment scores were computed by the mean number of fragments in DMRs. Then, fragment size selection was conducted to reduce the effect of plenty of non-tumor cfDNA in plasma based on that ctDNA fragments are shorter than non-tumor cfDNA fragments. Next, a fragment-based strategy was devised to statistically infer the origin (malignant or not) of each fragment, based on the distributions of DNA methylation pattern of tissues in DMRs. Finally, a predictive score (cfMeth score) based on cfDNA methylation ratio in each plasma sample was computed using a random forest classifier