2.1 An improved method was developed to obtain ctDNA with less background contamination and a high-quality WGBS library for sequencing

We developed an improved method to build a WGBS library in trace quantities of ctDNA. We utilized the selective binding properties of the silica membrane to extract ctDNA from plasma using QIAamp circulating nucleic acid kits (Figure 1A). First, the plasma samples were lysed in an optimized buffer and adjusted to binding conditions. Then, the samples were loaded directly onto a spin column. In this step, ctDNA was bound to the silica membrane, and contaminants were completely removed in wash steps. Finally, pure ctDNA was eluted in small volumes of a low-salt buffer for downstream applications. Meanwhile, we optimized the extraction method of ctDNA and found that higher amounts of ctDNA could be extracted by adding EDTA and proteinase K (Figure 1B). The explanation for the improvement in ctDNA extraction performance might be that the addition of EDTA could inhibit the degradation of ctDNA by nucleases in blood. In addition, proteinase K could improve ctDNA yield by releasing protein-bound ctDNA. Moreover, the ctDNA amount obtained was higher using the QIAamp circulating nucleic acid kit-based protocol than the conventional magnetic-bead-based protocol. The quality control analysis demonstrated that the DNA fragments have sharp peaks of approximately 160–180 bp (Figure 1C), indicating that our method could obtain high-quality ctDNA.

As shown in Figure 1D, purified ctDNA was ligated with an adapter and converted using bisulfate with an “all-in-one tube” operation. Specifically, the process of end repair, dA-tailing, adapter ligation, and bisulfite conversion was performed in the same tube, considering fewer ctDNA requirements and material losses. Then, the fragments were completed by PCR amplification, followed by purification. We optimized the purification method of the ctDNA library by comparing the agarose gel extraction method with the magnetic bead method. The head-to-head comparison showed that the magnetic bead method significantly increased the recovery ratio of the resulting library (Figure 1E). Bisulfite treatment of ctDNA led to the conversion of unmodified cytosines to uracil while keeping 5-methylcytosine unchanged, which could be mapped at a single base resolution after PCR and sequencing. With the above efforts, our ctDNA-WGBS method could be applied for whole-genome base-level resolution detection of 5-methylcytosine in ctDNA from breast cancer patients. Eventually, we could obtain the whole-genome DNA methylation landscape of breast cancer and reveal potential biomarkers.

2.2 ctDNA methylation biomarkers enabled sensitive detection of breast cancer

To validate the efficacy of the method, we collected breast cancer samples and healthy controls (Table S1). We first analyzed the genome-wide methylation patterns to compare the differences in the global methylation level and distribution of ctDNA from breast cancer patients and healthy controls. The results showed that the overall methylation level had no significant differences (Figure 2A). By dividing the methylation beta-level (0–1) into 10 intervals, the results revealed that the breast cancer patients and healthy controls showed similar distribution patterns (Figure 2B). Then, we tried to compare the ctDNA methylation data of breast cancer patients with healthy controls to screen candidate CpG sites. Feature selection was performed by least absolute shrinkage and selection operator (LASSO)-penalized logistic regression, resulting in 26 ctDNA methylation CpG sites as potential biomarkers for breast cancer diagnosis. The hierarchical clustering results suggested a precise classification of healthy individuals (n = 7) and patients (n = 74) in the training dataset using the methylation levels of these CpG sites (Figure 3A). The receiver operating characteristic (ROC) curve demonstrated that the sensitivity, specificity, and area under the curve (AUC) were 100%, 75%, and 0.889, respectively (Figure 3B). Meanwhile, the box plot shows the position of 26 potential biomarkers (Figure 3C).

2.3 ctDNA methylation biomarkers enabled the sensitive diagnosis of early-stage breast cancer

With these data in hand, we then decided to evaluate the early diagnostic ability of ctDNA methylation CpG sites. Specifically, we recruited seven early-stage breast cancer patients from the Tianjin Medical University Cancer Institute and Hospital (TMUCIH), and the healthy control groups enrolled age-matched Chinese women who were confirmed by physical examination. The heatmap results suggested a precise classification of healthy persons and early breast cancer patients (Figure 4A). Then, feature selection was performed by a random forest algorithm, resulting in 10 CpG sites as potential biomarkers for the early diagnosis of breast cancer (Figure 4B). The principal component analysis (PCA) data showed that healthy individuals and early-stage breast cancer patients were precisely classified (Figure 4C). Furthermore, we validated the identified biomarkers for early diagnosis using independent cohorts named test set 2, which comprised 12 Chinese female patients with early-stage breast cancer recruited from TMUCIH between May 2018 and October 2018. The ROC curve demonstrated that the AUC was 0.783, and the sensitivity and specificity were 93.44% and 50%, respectively (Figure 4D). Collectively, these results show that the 10 ctDNA methylation biomarkers obtained could sensitively distinguish early-stage breast cancer.

2.4 ctDNA methylation biomarkers distinguished different subtypes of breast cancer

The heterogeneity of breast cancer suggests that the tumor may consist of phenotypically different cancer cell populations with diverse properties and expressions of various functional molecules. According to hormone and epidermal growth factor receptor expression status, breast cancer can be divided into several molecular subtypes: triple-negative breast cancer (TNBC), luminal A (LA), luminal B (LB), human epidermal growth factor receptor 2 (HER2) enriched, normal-like, and so forth.

Encouraged by sensitive and specific early-stage breast cancer detection, we constructed three breast cancer datasets to validate the subtyping ability of differentially methylated CpG sites, including TNBC patients and non-TNBC controls, LA patients and non-LA controls, and LB patients and non-LB controls. For each dataset, we selected particular differentially methylated CpG sites for analysis. First, the heatmap plot showed that the obtained differentially methylated CpG sites could be used to divide the TNBC patients (n = 15) and non-TNBC controls (n = 30) into two groups (Figure 5A). Then, the corresponding CpG sites performed well for subtyping LA patients (n = 3) and non-LA controls (n = 22) (Figure 5B). Finally, the methylation level of CpG sites in ctDNA could also efficiently discriminate between LB patients (n = 11) and non-LB controls (n = 14) (Figure 5C). Collectively, these results suggested that the differentially methylated CpG sites we identified could be used to subtype breast cancer by heatmap analysis.

Based on the encouraging outcomes, we performed PCA to further assess the capability of the as-obtained ctDNA differentially methylated CpG sites to classify the subtypes of breast cancer. As shown in Figure 5D, the results indicated a precise classification of TNBC patients and non-TNBC controls. In addition, two distinct clusters of LA patients versus non-LA controls were separated (Figure 5E). Meanwhile, the PCA demonstrated that the LB patients were clustered together, and non-LB controls were grouped at the top of the coordinate plot (Figure 5F).

To further validate our obtained biomarkers, we conducted external validation of the identified CpG sites using independent The Cancer Genome Atlas (TCGA) dataset. The t-distributed stochastic neighbor embedding analysis indicated that TNBC (n = 15), LA (n = 12), HER2-positive (HER2⁺) LB (n = 10), HER2-negative (HER2^–) LB (n = 26), and HER2 (n = 9) patients could be well distinguished (Figure 5G). Although further systematic evaluations are still needed, these preliminary results indicate that our study provides an unprecedented method to develop novel biomarkers complementary to traditional diagnostics for classifying various subtypes of breast cancer.

4.1 Sample collection

We recruited two cohorts of Chinese breast cancer patients from the Tianjin Medical University Cancer Institute and Hospital (TMUCIH): (i) Test set 1 consisted of early-stage (n = 7) and advanced-stage (n = 44) breast cancer female patients of Chinese descent from the TMUCIH between December 2016 and December 2017. In sum, we enrolled 51 Chinese female patients with early/advanced breast cancer (50.87 ± 11.05 as the mean age ± standard deviation [SD]) in test set 1. (ii) Test set 2 comprised Chinese female patients with early-stage (n = 12) and advanced-stage (n = 11) breast cancer recruited from the TMUCIH between May 2018 and October 2018. For test set 2, 23 Chinese female patients with early/advanced breast cancer (53.47 ± 8.88 as the mean age ± SD) were recruited. Collectively, there were 74 female patients with breast cancer enrolled in this study.

The healthy control groups enrolled age-matched Chinese women without cancer from the Beijing Institute of Genomics (BIG), confirmed by physical examination, ultrasound scans of breast, and mammographic screening. Informed consent or a waiver of consent was obtained from each participant. In total, seven healthy women were recruited into healthy control groups.

4.2 Phenotypic evaluation

The diagnosis of breast cancer was histologically confirmed. We conducted phenotypic analyses of age, menopausal status, stage, pathology features, and molecular subtype in the two cohorts. Information on estrogen receptor (ER) and progesterone receptor (PR) status was assessed by immunohistochemistry (IHC) and extracted from pathology reports. Specifically, the paraffin samples were cut into 4-μ e paraffin samples were cut into hology features, and molecular subtype in the two cohorts. Information on estrogen receptor (E mol/L citrate buffer (pH = 6.0) for 10 min for antigen retrieval. After cooling, the slides were washed with PBS (pH = 7.4) and incubated with 0.3% H₂O₂ solution for 10 min to block endogenous peroxidase activity. Primary polyclonal antibodies against ER, PR, and HER2 were applied to the samples and incubated overnight at 4al r subtype in the two cohorts. Informath PBS, and the antibody reaction was visualized using a fresh substrate solution containing diaminobenzidine. Images were captured by a Leica DC500 camera (Solms, Germany) on a microscope equipped with Leica DMRA2 fluorescent optics under 10 orts. Innification, and the expression level was defined as the mean density of ER-, PR-, and HER2-specific staining/area. Tumors were classified as ER/PR-negative if IHC staining of tumor cell nuclei was less than 1% reactivity. HER2 was assessed through IHC or fluorescence in situ hybridization (FISH) if the IHC results were borderline. The clinical grouping of molecular subtypes was defined by the status of hormone receptor and HER2 according to the St. Gallen 2017 criteria.³⁶

4.3 Circulating tumor DNA extraction

Whole blood of healthy and cancer samples was collected using ctDNA blood collection tubes (Strect, USA). Plasma was then separated by centrifugation at 1900 × g for 10 min and 18,000× g for 10 min at room temperature with EDTA and protinease K. ctDNA was extracted from plasma using the QIAamp Circulating Nucleic Acid Kit (55114) or magnetic beads method according to the manufacturer's instructions. The ctDNA concentration was determined using a Qubit dsDNA HS Assay Kit (Life Technologies, Q32851). The extracted ctDNA was stored at −80°C for further library construction.

4.4 ctDNA methylation library construction

As shown in Figure 1D, we used the DNA to build the methylation library. We mixed 20 ng ctDNA and 1/1000–5/1000 ng unmethylated lambda DNA (D1521, Promega). Unmethylated lambda DNA was sheared by a Covaris S220 instrument (Life Technologies, Thermo Fisher Scientific), and the size of fragments was approximately 200 bp. First, DNA fragments were synthesized via end-repair and A-tailing reactions. Second, NEBNext Ultra Ligation Module (E7445S/L, NEB) reagent was used to ligate the methylation adaptor. Then, we used the EZ DNA Methylation-Gold Kit (D5005, Zymo) to bisulfite-converted DNA according to the manufacturer's recommendations. Next, DNA was amplified by PCR, which was carried out using primers. The primers came from Illumina next-generation sequencing. PCR was performed to amplify the fragments. DNA gel extraction and magnetic bead extraction were performed to compare the different influences of library quality. Finally, we selected effective captures (∼290 bp) as the resulting library using Agencourt AMPure XP beads (Beckman Coulter). All binding and washing procedures were performed at room temperature. The purity of the libraries was analyzed by Qubit dsDNA HS assay (Thermo Fisher). A total of 100 G data were obtained for each sample.

4.5 Quality control and mapping of WGBS data

The raw sequencing reads (fastq format) were trimmed to remove sequencing adapters, amplification primers, and low-quality bases in read ends using trim_galore (version 0.4.2) (http://www.bioinformatics.babraham.ac.uk/projects/trim_galore). After quality control, the reads were mapped to the human reference genome (hg19) using Bismark (version 0.14.2) with default parameters.³⁷ Next, the PCR duplicates were also removed using Bismark (version 0.14.2) with default parameters.³⁷ The bisulfite conversion rate was calculated by the spike-in of totally unmethylated lambda DNA. Only the mapped and duplicate-removed reads (bam format) were used for the subsequent bioinformatics analysis.

4.6 Data processing and analysis

We sequenced DNA methylation libraries by using X-ten systems (Illumina). The total reads were assessed by FASTQC (version 0.11.8). The raw sequencing reads were then cleaned by Trim Galore (version 0.4.5) to clip sequencing adapters and low-quality reads. Subsequently, the trimmed reads were aligned against the human reference genome using the Bismark (version 0.14.3) tool. Finally, differential methylation analysis of the DNA methylation sequencing data was performed by the limma package. For each dataset, the top CpG sites with p-value <0.05 and absolute mean methylation difference >0.15 were selected. Sites were annotated by PeakAnnotator (version 1.4).