2.1 Study populations and samples

TNBC fresh-frozen tissues from 123 patients and plasma samples from 139 patients (60 patients provided both tumor tissue and plasma) participating in the Städtisches Klinikum Karlsruhe Deutsches Krebsforschungszentrum Breast Cancer Study (SKKDKFZS) were selected.¹¹ Information on known and potential risk factors and follow-up was collected for all study participants from medical records, pathology reports and/or questionnaires. Table S1 summarizes the main characteristics of the study participants including the established BC risk factors age at diagnosis, menopausal status (premenopausal, postmenopausal), family history of breast/ovarian cancer (yes, no), body mass index (BMI) (<20, 20 to <25, 25 to <30, ≥30 kg/m²) and parity (yes, no). Table 1 shows the histopathological data of the TNBC patients including histology (ductal, lobular, ductolobular, medullary, other), histological grade (G1, G2, G3), tumor size (in situ, T1, T2, T3, T4), lymph node status (N0, ≥N1), metastatic status (M0, M1), stage (0, 1, 2, 3, 4) and follow-up data (number of recurrences, number of deaths; median durations of follow-up and recurrence-free time [years]).

DNA was isolated from 123 fresh-frozen tumor samples of TNBC cases. After quality control, the DNA samples were applied to down-stream DNA methylation analysis. The plasma samples were derived from 139 TNBC patients and 84 healthy controls. Of the TNBC plasma cohort, 60 patients provided both tumor and plasma (test set) and 79 patients provided plasma only (validation set). To independently confirm our data from the tumor-informed approach, we analyzed the plasma samples of the two sample sets separately.

Blood samples from healthy individuals were collected by the German Red Cross Blood Service of Baden-Württemberg-Hessen (Mannheim, Germany) from September to December 2019 for plasma isolation. All blood donors were females aged 30 to 55 years at the time of blood donation. Details on the study populations can be found in Methods S1.

2.2 Genome-wide DNA methylation profiling

DNA methylation profiling of 123 TNBC tissues was performed at the DKFZ Genomics and Proteomics Core Facility (Heidelberg, Germany) using the Illumina Infinium HumanMethylation450K BeadChip (n = 52) and the EPIC array (n = 71) according to the manufacturer's instructions. To account for cell type heterogeneity, we used the Houseman algorithm for reference-based deconvolution using DNA methylation data from sorted cell types. Estimated proportions of various mammary cells (fibroblasts, epithelial and endothelial cells), adipocytes and immune cells (using the LUMP: leukocytes unmethylation for purity algorithm) were used as covariates to account for cell type heterogeneity in bulk tumor methylation data.^{12, 13}

Additionally, to minimize the false positive detection and to increase the signal-to-noise ratio in cfDNA analysis, we selected CpG sites that were almost completely methylated (β-value >0.9) or unmethylated (β-value <0.1) in peripheral blood DNA (Figure 1) using methylation data from our previous genome-wide methylation study on 231 TNBC cases and 231 controls.¹⁴

Next, to identify DMRs, we compared the methylation level at the prioritized CpG sites (n = 84 781) between the two sets of TNBC tissues (first set: n = 52, second set: n = 71) and normal breast tissues (n = 221) (Figure 1). The methylation datasets under evaluation are provided in Table 2.

2.3 Verification and external validation of the identified DMRs

The identified DMRs were verified by amplicon bisulfite sequencing at single-molecule resolution. PCR primers were designed manually and also using the AmpliconDesign webserver to amplify candidate loci from bisulfite-converted DNA¹⁵ (Table S2). For an external validation, DNA methylation datasets of TNBC tissues (n = 73) and normal breast tissues (n = 98) from The Cancer Genome Atlas (TCGA-BRCA) were applied. Details on the method can be found in Methods S1.

2.4 Selection of top classifier DMRs by machine learning

To select CpG sites with the most predictive power, an eXtreme Gradient Boosting (XGBoost) model was trained based on methylation levels of CpG sites overlapping with the selected DMRs. In this respect, the TNBC and normal breast tissue cohorts were split into training (n = 247; Healthy = 159; TNBC = 88), validation (n = 61; Healthy = 39; TNBC = 22) and holdout (n = 36; Healthy = 23; TNBC = 13) datasets. Model training and hyperparameter tuning was performed on the training cohort using a fivefold cross-validation approach with a grid-search algorithm using the R caret package. Details on the method can be found in Methods S1.

2.5 Isolation of circulating cell-free DNA and bisulfite conversion

Plasma samples of 139 TNBC patients (test set: n = 60, validation set: n = 79) and 84 healthy controls (test set: n = 36, validation set: n = 48) were kept at −80°C until use. For each sample, cfDNA was extracted from 2 mL of plasma using the QIAamp MinElute ccfDNA Kit (Qiagen, Hilden, Germany). Bisulfite conversion was carried out on cfDNA samples using the manufacturer's protocol for low concentration DNA samples (EpiTect bisulfite kit; Qiagen).

2.6 MethyLight droplet digital PCR

For investigation of the top DMRs in circulating cfDNA, MethyLight droplet digital PCR (ddPCR) was performed as previously described by us.¹⁴ Specific TaqMan probe assays were designed and ddPCR was employed for sensitive detection of aberrant methylation at target DMRs in cfDNA samples. The TaqMan probe assays were designed to include multiple nearby CpG sites considering co-methylation pattern of those CpGs. The length of amplicons was kept as short as possible to maintain higher likelihood of detection. All steps including droplet generation, thermal cycling and droplet reading were performed according to the manufacturer's protocols (Bio-Rad). The list of primer and Taqman probe sequences is provided in Table S3.

2.7 Statistical analyses

The diagnostic performance of individual selected CpGs was assessed using logistic regression and receiver operating characteristic (ROC)/area under the curve (AUC) analysis. In TNBC cases, the association of CpG methylation levels with clinical and epidemiological factors (age, menopausal status, family history of breast/ovarian cancer, BMI, parity), histopathological tumor parameters (grade, size, lymph node status, metastatic status, stage) and overall survival (OS) and recurrence-free survival (RFS) was assessed. OS was defined as the time between TNBC diagnosis and death or last follow-up, whichever occurred first. Recurrence-free survival was defined as the time between TNBC diagnosis and reappearance of disease (locoregional relapse, distant metastasis). Mann-Whitney test, Jonckheere-Terpstra trend test and Spearman's correlation coefficient were used to assess associations between methylation levels and clinical, epidemiological and histopathological parameters. The impact of methylation level on OS was analyzed in a Cox regression model. Kaplan-Meier estimates and log-rank test were derived for methylation levels at median cut-off. To account for established prognostic factors, a multivariable Cox regression model including age, tumor grade, stage, tumor size and node status was fitted. Individual P-values were adjusted for multiple testing using Holm correction to control the family-wise error rate. All analyses have been done using R 3.6 with add-on packages rms, survival and pROC.

3.1 Patient and sample characteristics and data sets analyzed

We analyzed genome-wide methylation data (Illumina 450 k and EPIC arrays) of a total of 852 DNA samples from peripheral blood of 231 TNBC patients and 231 controls, 123 TNBC and 221 normal breast tissues and 46 reference cell including neutrophils, B cells, CD4+ T cells, CD8+ T cells, monocytes, natural killer cells and various mammary cells (Table 2, Figure 1). Principal component analysis (PCA) using methylome data showed a clear separation of the samples into different clusters (Figure 2A). The TNBC tissues formed a more dispersed cluster that may be due to tumor heterogeneity and different ratios of tumor to nontumor cells in each sample. Clinical, epidemiological and histopathological characteristics of the TNBC patients and their tumors as well as follow-up data are provided in Tables S1 and 1, respectively. In summary, family history of breast/ovarian cancer was reported by 11% of patients. Ductal invasive carcinoma was the most common histological type, followed by lobular carcinoma. The majority of patients who provided tumor tissue and/or plasma had nonmetastatic disease (93% and 96.4%, respectively) and only a few patients had metastatic disease at the time of diagnosis and sampling. The mean and median ages at diagnosis of TNBC were 57.0 and 58.6 years, respectively. Median follow-up of TNBC patients was 6.5 years for OS and 2.1 years for RFS.

3.2 Identification, verification and validation of differentially methylated regions

The data analyses performed to identify the most promising DMRs are shown in Figure 1. Considering that PBLs are the main source of cfDNA in healthy individuals, we prioritized those CpG sites that were almost completely methylated (β-value >0.9) or unmethylated (β-value <0.1) in peripheral blood DNA (Figure 2B) using data from our previous study.¹⁴ Following differential analysis, 23 DMRs (18 hypermethylated, 5 hypomethylated) with 52 CpGs were identified. The heatmap of hierarchical clustering shows the TNBC tissues and the normal breast tissues grouped into two main clusters (Figure 2C). The probes with their genomic location; position relative to known genes and CpG islands, and methylation data at the 52 CpG sites are shown in Table S4. The identified DMRs yielded AUCs ranging from 0.739 to 0.987 for discrimination of TNBC from normal breast tissue. Methylation levels at all 52 CpG sites were associated with TNBC in unadjusted and age-adjusted conditional logistic regression analysis (all P < .001; Table S4).

Using amplicon bisulfite sequencing, all DMRs were verified. The coverage and quality statistics of the targeted bisulfite sequencing analysis are for each sample summarized in Table S5. The methylation status of all DMRs was consistent with the initial data from the Illumina methylation arrays. Representative results of the amplicon sequencing of the 23 DMRs in TNBC and normal breast tissues are shown in Figure S4.

For independent validation of the DMRs, we analyzed the available TCGA-BRCA data. Statistically significant methylation differences between the TNBC and normal breast tissues were found, with a P-value ≤.0001 for all available CpG sites and methylation differences in the range of 0.22 to 0.43, confirming our findings (Figure S2).

Aberrant DNA methylation could affect expression of the associated genes, which may have prognostic potential. In that respect and using the Kaplan-Meier (KM) plotter tool, we analyzed gene-expression data of several combined BC studies to see whether expression of genes associated with the DMRs predicts patient survival.¹⁶ Using the median to split gene expression level, expression levels of nine genes including CARD11, EPSTI1, IFFO1, KDELR2, MIAT, PPP1R16B, SLC7A4, TXNR and ZNF750 were associated with OS of BC patients (Figure S3). The methylation levels of three of these genes, CARD11, IFFO1 and KDELR2, were also associated with OS in our study cohort (Figure 3A).

3.3 Correlation of methylation levels with clinical, epidemiological and histopathological parameters

Next, we assessed, whether methylation levels of 52 CpG sites which are located in the 23 selected DMRs are associated with age, menopausal status, family history of breast/ovarian cancer, BMI, parity and histopathological tumor parameters including histological grade, tumor size, lymph node status, metastatic status and stage. Two statistically significant associations were detected in analysis adjusted for multiple testing. Methylation levels at six CpG sites in three DMRs corresponding to the PPP1R16B, SFRP5 and CPXM1 genes were associated with age at diagnosis (Spearman's rank correlation, P_adj < .05) (Table S6). Patients with an early age at diagnosis had lower methylation levels than those with a late age. Further, methylation levels at 10 DMRs corresponding to the ALPL, PPP1R16B, FAM110B/IMPAD1, SFRP5, KIAA1949, PRDM16, CPXM1, TXNRD1, SPHK2 and PROCA1 genes were associated with tumor grade (Mann-Whitney test, P_adj < .05) (Table S7). Methylation levels were significantly different in grade 3 tumors compared to grade 1/grade 2 tumors (Figure S4).

3.4 Prioritization of top classifiers for liquid biopsy applications

A machine learning algorithm was applied to select CpG sites within the DMRs that have the highest predictive power to discriminate TNBC patients from healthy controls. We, therefore, trained an XGBoost model based on the training cohort that predicted patients of the validation and holdout datasets independently with an accuracy of 100%. Subsequent feature selection based on the CpG sites with the highest model contribution, revealed 12 CpGs that accurately discriminate TNBC patients from healthy controls (Figure S5A,B). With respect to diagnostics, the AUC values of all selected top sites were close to one (range 0.94-0.99), demonstrating their high potential as biomarkers for TNBC diagnosis (Figure 3B). Somewhat lower AUC values (range 0.69-0.93) were calculated using TCGA-BRCA data (Figure 3C).

3.5 Sensitive detection of methylation markers from cfDNA

Due to the limitations within the amount of cfDNA for the analysis of all 12 CpGs, the top 7 CpGs (6 DMRs) from this selection were used to further analyze plasma-derived cfDNA samples from TNBC cases and controls. The six designed TaqMan assays were able to specifically detect the desired alleles in a bulk of background unmethylated/methylated DNA. Four assays were specific for the methylated allele and two for the unmethylated alleles. The copy numbers per microliter of each top six DMR candidates was measured on cfDNA samples from TNBC patients and healthy controls using methylation-specific ddPCR. Copy numbers at three DMRs (for the methylated allele of SPAG6, and for the unmethylated alleles of LINC10606 and TBCD/ZNF750) were significantly higher in TNBC patients compared to controls in both the test and the validation set (Figure 4A,B). The AUC values to discriminate TNBC cases from healthy controls were 0.68, 0.70 and 0.73 for the top three DMRs The sum of the copy numbers of the three significant DMRs was calculated and used as methylation score for each sample. The median methylation score was 2.14 for TNBC patients and 0.86 for healthy individuals. The difference in methylation scores between cases and controls in both sample sets was highly significant (P < .0001; Figure 4C,D). ROC analyses yielded AUC values of 0.78 on the test set and 0.74 on the validation set (Figure 4E,F). In addition, we also compared the methylation score in patients with different disease stages. Methylation scores in TNBC patients with late-stage disease (stages 3 and 4) were slightly higher compared to those with early-stage disease (stages 0 and 1 and stage 2), but the differences were not statistically significant (Figure S6).