MammaTrace – a tumor naïve cell-free DNA methylation test for minimal residual disease in breast cancer

Abstract

Metastatic breast cancer (MBC) remains an incurable disease with a 5-year overall survival below 25%. Metastases often emerge from subclinical, disseminated tumor cells that persist after primary therapy, referred to as minimal residual disease (MRD). Detecting MRD is critical for identifying patients at high risk of recurrence and enabling timely intervention. We constructed MammaTrace, a plasma-only, tumor-naive cfDNA methylation assay for MRD surveillance in breast cancer. Candidate differentially methylated regions (DMRs) associated with metastasis and recurrence were identified directly from plasma cfDNA using whole genome bisulfite sequencing (WGBS) of MBC, healthy controls, and treatment-naive patients with stage I–III breast cancer who either developed recurrence or remained disease-free. These DMRs were incorporated into a targeted capture panel and used to train a binary classification model. The assay was evaluated in an independent longitudinal cohort of 107 early-stage breast cancer (EBC) patients treated with curative intent, with serial plasma samples collected prior to recurrence. MammaTrace demonstrated strong performance in distinguishing MBC from healthy samples, achieving a ROC AUC of 0.96. For MRD detection in EBC samples collected more than 9 months after surgery, MammaTrace achieved a sensitivity of 91% and specificity of 83% (AUC = 0.88). A positive MammaTrace score preceded clinically confirmed recurrence by a median of 457 days, providing substantial lead time for therapeutic intervention. MammaTrace enables detection of MRD in breast cancer patients, providing substantial lead time prior to clinical recurrence. This approach may improve risk stratification and support earlier therapeutic intervention to delay or prevent metastatic progression.

Introduction

Breast cancer remains the leading cause of cancer-related death among women worldwide¹. While approximately 25% of metastatic breast cancer (MBC) presents as de novo disease, the majority of cases arise following an initial diagnosis of localized early-stage breast cancer (EBC) treated systemically with curative intent². Detection of minimal residual disease (MRD), also known as molecular residual disease, defined as the presence of subclinical, disseminated tumor cells or tumor-derived molecular signals persisting after primary and adjuvant therapy, is strongly associated with an increased risk of distant recurrence and progression to MBC³.

Although MRD testing is standard of care in hematological malignancies, its application in solid tumors, including breast cancer, is still evolving, with no fixed time points or standardized testing protocols. This is further reflected in the current National Comprehensive Cancer Network (NCCN) Breast Cancer Guidelines, which do not include recommendations for MRD surveillance in routine clinical practice⁴. Recent advancements have established MRD surveillance as a valuable tool for monitoring treatment response, predicting relapse, and guiding therapeutic decisions in solid tumor types^5–10. MRD surveillance enables detection of trace amounts of residual disease after treatment and may provide earlier indication of recurrence than conventional clinical or radiographic methods. However, many MRD studies in solid tumors have focused on early post-treatment risk assessment, typically within the immediate post-operative period (1–3 months after surgery), to guide adjuvant therapy decisions^11–13. In contrast, the role of MRD testing in long-term surveillance remains poorly defined, despite the risk of breast cancer recurrence years after completion of primary therapy. Long-term studies evaluating the clinical utility of commercially available MRD tests beyond the immediate post-treatment time frames remain limited, leaving a critical gap in understanding how MRD may be used for extended surveillance. Despite advances in risk stratification and increasingly effective systemic and locoregional therapies, approximately 20–30% of EBC patients ultimately develop metastatic disease¹⁴, underscoring the need for sensitive assays capable of detecting MRD during extended follow-up.

The development of circulating tumor DNA (ctDNA) assays has enabled blood-based monitoring of MRD in recent years. Peripheral blood contains multiple tumor-derived analytes, including circulating tumor cells, microRNAs, extracellular vesicles, and ctDNA, which harbor tumor-specific genetic and epigenetic alterations^15–20. There are several advantages to developing liquid biopsy assays based on aberrant DNA methylation over those detecting other molecular alterations such as point mutations or serum-based protein markers^21–27. DNA methylation changes occur more frequently across the genome than single nucleotide variants or copy number alterations, potentially improving analytical sensitivity^25,28. Additionally, DNA methylation-based assays target multiple altered regions with numerous CpG sites, allowing for a lower limit of detection compared to protein-based markers or DNA mutations²⁹. Furthermore, since DNA methylation is altered in tumor cells, these changes can be detected in the presence of background cfDNA derived from normal cells^24,29.

MRD detection strategies using ctDNA can be broadly classified as tumor-informed or tumor-naive. Tumor-informed assays begin with sequencing tumor tissue to identify patient-specific genomic alterations that can be subsequently monitored in plasma using personalized panels. Since they track known variants that arise in the patient’s tumor, these approaches typically achieve very high specificity and high sensitivity for detecting low tumor fractions. However, they require access to tissue and involve greater logistical complexity, longer turnaround times, and higher costs associated with tumor sequencing and bespoke assay design. In contrast, tumor-naive (also known as ‘tumor-uninformed’ or ‘tumor-agnostic’) approaches do not require tumor tissue or prior knowledge of patient-specific tumor alterations⁸. Instead, they detect MRD using fixed, non-patient-specific panels that target cancer-specific alterations shared across multiple tumors of a given cancer type^30,31. Tumor-naive approaches offer broader applicability, simplified workflows, and reduced cost, but historically have demonstrated lower sensitivity for detecting very low tumor fractions compared with tumor-informed assays³⁰.

In the current study, we aimed to develop an MRD surveillance assay for breast cancer specifically designed for longitudinal monitoring beyond the immediate post-operative window. Rather than focusing solely on identifying markers of tumor burden, we prioritized cfDNA methylation markers specifically associated with breast cancer recurrence and metastasis. Here, we introduce MammaTrace, a novel, tumor-naive, plasma-only cfDNA methylation-based liquid biopsy assay designed for MRD surveillance and recurrence prediction in breast cancer. While MammaTrace does not rely on prior knowledge of the primary tumor, it is uniquely informed by molecular markers associated with metastatic disease, allowing for sensitive detection across diverse stages of cancer progression. This strategic focus on metastatic-associated markers acknowledges that disseminated breast cancer cells may persist for years or decades without progressing to clinically evident metastatic disease, suggesting that not all detectable MRD predicts recurrence. Focusing on cfDNA markers associated with recurrence risk rather than tumor burden alone may therefore increase the accuracy of the assay in predicting recurrence. Furthermore, our plasma-only feature discovery approach captures systemic, biologically relevant cfDNA signals directly from blood, enhancing clinical applicability in settings where tumor tissue is unavailable and better reflecting the dynamic processes underlying cancer progression. Moreover, whether and how mechanisms of tumor dormancy may influence MRD is also yet to be determined³².

MammaTrace was evaluated in a longitudinal cohort of 107 patients with EBC treated with neoadjuvant chemotherapy followed by surgery. MammaTrace correctly identified MRD in 10 of 11 patients who subsequently developed metastatic disease (sensitivity = 0.91 at 0.83 specificity), with a median lead time of 15 months prior to clinical recurrence. These findings demonstrate that MammaTrace is a highly accurate, novel cfDNA methylation-based MRD assay with potential clinical utility for identifying EBC patients who remain at elevated risk of recurrence despite standard curative therapy.

Results

cfDNA methylation differences associated with breast cancer metastasis

To identify cfDNA methylation alterations associated with MBC for use in developing an MRD assay, we employed a two-pronged strategy. This approach aimed to identify DMRs linked to distant recurrence by analyzing both diagnostic EBC samples collected prior to recurrence and MBC samples collected after recurrence. For this, we performed WGBS on 6 cfDNA pools, including 5 MBC pools representing each molecular subtype, plus a separate AA TNBC pool. Two healthy female pools from predominantly NHW women and one pool from healthy AA women were used as controls. In total, the 5 MBC cfDNA patient pools represented 41 individuals, the AA TNBC pool represented 21 individuals, and the 3 healthy cfDNA pools represented 39 individuals. A breakdown of the number of patients in each pool can be found in Table S1. DMRs were identified between subtype-specific MBC cfDNA pools and healthy controls (Fig. 1, Table S1). AA TNBC samples were compared only to the AA healthy pool to control for race-specific methylation alterations. We identified a total of 2353 DMRs from MBC pools versus healthy control pools — specifically, HR + /HER2- N = 270 DMRs, HR-/HER2 + N = 362 DMRs, HR + /HER2 + N = 507 DMRs, and TNBC N = 1,214 DMRs (401 TNBC DMRs were derived from the AA-specific TNBC pool). The DMRs identified for each subtype were largely distinct (Fig. 2A). We also examined the overlap between DMRs identified from cfDNA in predominantly NHW TN MBC and AA TN MBC patients. Only 4 DMRs overlapped between the 813 NHW TNBC-associated DMRs and the 397 AA TNBC-associated DMRs (Fig. 2B); however, these patients were not stage-matched. These findings highlight the impact of race on cfDNA methylation changes in TNBC and possibly other breast cancer subtypes. Hierarchical clustering and uniform manifold approximation and projection (UMAP), based on beta values in MBC DMRs, showed clear separation of the MBC sample pools from the healthy control pools (Fig. 2C,D).

Fig. 1. Patient sample summary and study workflow.

A total of 331 unique patient donors were used in this study. The study used plasma from patients with metastatic breast cancer (MBC), metastatic and non-metastatic African American patients with TNBC (AA TNBC), healthy individuals, and patients with early-stage breast cancer (EBC) who were undergoing neoadjuvant chemotherapy. Patients with EBC had multiple collections over time. The pooled MBC/Healthy cohort and unpooled EBC at diagnosis cohort was used for WGBS feature discovery of MRD markers. The number of plasma samples used for feature discovery is shown in the left panel. The right panel shows the number of plasma samples used for targeted validation of MRD markers and model construction. The number of plasma samples overlapping between feature discovery and targeted validation are indicated by a superscript^1–4. The number of non-MBC African American (AA) TNBC plasma samples are also indicated by a superscript⁵.

Fig. 2. Feature discovery for MRD markers from cfDNA WGBS analysis.

A Venn diagrams showing overlap of DMRs identified in subtype-specific MBC cfDNA pools. B Venn diagram showing overlap of DMRs between NHW and AA TNBC cfDNA pools. C, D Heatmap and UMAP based on mean beta values across all MBC-associated DMRs, demonstrating separation between healthy and MBC pools used for feature discovery. UMAP point colors and the heatmap annotation bar denote healthy and MBC samples. E, F Heatmap and UMAP based on mean beta values across all early-stage breast cancer associated DMRs, showing separation by disease status, including DFS, REC-D, and REC-LR patients.

Next, we performed WGBS on 21 baseline samples from the EBC cohort (Fig. 1). Specifically, we analyzed the baseline samples from patients at diagnosis before they developed distant (REC-D, N = 10) or locoregional recurrence (REC-LR, N = 2), and we also analyzed the baseline samples from patients at diagnosis who remained disease-free for greater than 5 years (N = 9) (Table S2). These baseline samples comprised all major breast cancer molecular subtypes – TNBC (N = 14), HR-/HER2+ (N = 5), HR + /HER2- (N = 1), and HR + /HER2+ (N = 1) (Table S2). We performed DMR analysis to compare both local and distant recurrence patients to those who remained recurrence-free, and to the two non-AA healthy control pools. We found 2338 significant DMRs in REC samples when compared to disease-free and healthy samples. Hierarchical clustering/UMAP, based on recurrence status, showed a clear separation between the REC and disease-free samples (Fig. 2E, F).

Validation of DMRs and construction of the MammaTrace model for MRD detection in MBC samples using hybridization probe capture

In this study, we initially pooled cfDNA from MBC samples for WGBS, which — while comprehensive — is resource-intensive. Pooling offered several advantages: it reduced cost, enhanced the detection of common biologically relevant methylation changes across a patient group, improved the signal-to-noise ratio, and optimized resource use, particularly in the context of limited cfDNA analyte availability³³. We later analyzed unpooled baseline EBC samples from patients who either remained cancer-free for at least five years or eventually developed MBC. From this analysis, we prioritized 5,077 DMRs associated with metastasis, as described in the Materials and Methods. Pools were sequenced to a depth that was probabilistically calculated to ensure that each pooled sample was more likely than not to contribute at least one sequencing read across target regions, thereby preserving representation within the pooled data. The DMRs identified from pooled WGBS were subsequently evaluated at the individual-sample level using targeted hybridization capture, demonstrating that selected features were detectable across individual patients.

As a first step in evaluating these regions, we developed a targeted hybridization probe capture assay based on the 5077 DMRs to assess its ability to distinguish MBC from healthy controls. Using this probe panel, we performed hybridization capture sequencing on cfDNA from 72 MBC and 83 healthy individuals (total N = 155). The MBC samples included 34 HR + /HER2-, 16 HR-/HER2 + , 9 HR + /HER2+, and 13 TNBC patients (Figure S1, Table S3). Of these, 20 MBC and 12 healthy samples were included in pooled WGBS for DMR discovery (Fig. 1). The cfDNA validation cohort included 43 AA, 15 Asian, 24 HL, and 71 NHW individuals (Figure S1, Table S3). Additionally, two samples were sequenced across seven separate runs and served as technical replicates throughout the study. These controls showed highly consistent beta values at the target loci across all runs, demonstrating excellent reproducibility and the absence of batch effects (Figure S2). These findings are consistent with our previous work showing that hybridization probe capture for cfDNA methylation analysis is both highly reproducible and accurate^34,35.

To develop a classifier that distinguishes MBC from healthy control cfDNA samples, we employed a repeated, nested 5-fold cross-validation strategy: an outer loop to estimate generalization performance and an inner loop to optimize hyperparameters^6,36,37. Each inner and each outer CV was repeated 5 times. A schematic of this framework is shown in Supplementary Figure S1 of our previous study³⁵. A receiver operating characteristic (ROC) curve, based on the model’s prediction score, demonstrated excellent discrimination, with a ROC area under the curve (AUC) of 0.956 (95% CI: 0.925-0.988) (Fig. 3A, B). Performance was consistent across molecular subtypes with slightly lower performance in TNBC (AUCs: HR-/HER2 + = 0.983, HR + /HER2- = 0.972, HR + /HER2 + = 0.953, TNBC = 0.887) (Fig. 3C). Classifier performance also remained robust across racial and ethnic groups (AUCs: AA = 0.890, Asian = 0.990, HL = 0.994, NHW = 0.960) (Fig. 3D). A final model was then trained using all available MBC and healthy control samples without outer cross-validation. The output from this model, referred to as the MammaTrace score, was used to predict MRD status in the EBC cohort.

Fig. 3. Targeted validation and construction of the MammaTrace model.

A ROC curve showing cross-validated performance of the MammaTrace model in distinguishing MBC from healthy cfDNA samples. AUC with 95% confidence interval is shown. B Boxplot of cross-validated prediction scores in an independent cohort of MBC and healthy samples not used for feature discovery. Each point represents an individual sample. C ROC curves from the same cross-validated model in (A), stratified by MBC molecular subtype. The number of MBC patients per subtype, along with the number of healthy controls, is indicated. D ROC curves from the same cross-validation, stratified by race. The number of MBC and healthy samples per group and corresponding AUC values are shown. Five non-MBC AA TNBC patients were included in this analysis.

These findings demonstrate that methylation changes associated with MBC, identified using WGBS from samples collected prior to recurrence (baseline EBC) or after recurrence (MBC), are accurate, reproducible, and generalizable across patient subtypes and populations. This reinforces the clinical relevance of the identified DMRs and supports their use in building a model to distinguish MBC from healthy individuals. The resulting model lays the foundation for MRD detection in EBC patients.

Detection of MRD in early-stage breast cancer patients using MammaTrace

To evaluate the performance of MammaTrace for MRD detection in EBC patients selected for neoadjuvant chemotherapy and surgery, we applied our custom hybrid probe capture assay to 279 cfDNA samples collected longitudinally (Table S10) from 107 patients with EBC (Figs. 1, 4, Table S4). MammaTrace was initially assessed at each of the five clinical timepoints (baseline at diagnosis, post-neo, post-op, follow-up (9 months – 27 months), and post-rec, Fig. 4), regardless of whether patients had collections across multiple timepoints. This initial analysis aimed to identify the most informative timepoint for MRD detection and recurrence prediction relative to patient outcomes.

Fig. 4. Early-Stage Breast Cancer Cohort Description.

A Overview of blood collection time points for the EBC Mayo Clinic cohort. Blood was collected longitudinally from EBC patients (from 05/2015 to 10/2019) at diagnosis (baseline), post-neoadjuvant treatment (post-neo), within 9 months post-surgery (post-op) and at follow-up clinic visit (follow-up). The median follow up time was 12.4 months (range: 9 months – 27 months). A total of 107 patients with EBC were enrolled totaling 355 plasma samples collected longitudinally. Two patients had a recurrence at one of the designated blood draws and were designated “post-recur”. B A summary of each patient’s treatment timeline and serial blood collection grouped by breast cancer subtype (annotated in the left bar). The timelines are centered at the date of surgical resection for each patient, which is represented by the red vertical line (time = 0). Each colored dot represents one blood collection – the color denotes the timepoint of collection (baseline, post-neo, post-op, follow-up, and post-recur). The green vertical line represents the follow-up timepoint for the study. Open circles indicate samples that were collected but excluded from analysis due to QC failure. Samples not collected were not shown. C Not all patients had blood collections at every timepoint. The Venn diagram illustrates the number of patients with blood collections at every timepoint and their overlap. Post-recur collections (n = 2) were excluded from the diagram.

In DFS patients (those without recurrence), we observed an overall increase in the MammaTrace score at the post-neo and post-op timepoints compared to baseline, with average scores returning to baseline during follow-up. Specifically, the median MammaTrace score increased from 0.74 at baseline to 0.96 post-neo (Fig. 5A,B). In patients with matched samples (N = 43), MammaTrace scores increased by a median ΔMammaTrace score of 0.171 at the post-neo timepoint (Wilcoxon’s signed rank test p-value = 0.00097) (Figure S3). These findings suggest that cfDNA methylation patterns are altered following treatment, in contrast to ctDNA mutations, which have been shown to clear during and after chemotherapy^38–43. For DFS samples, MammaTrace scores subsequently decreased at the post-op and follow-up timepoints, with a median score of 0.90 at post-op (N = 48), and a median score of 0.75 at follow-up (N = 40), effectively returning to the baseline pre-treatment median score (Fig. 5A,B). This trend was also observed in paired samples, with a median ΔMammaTrace score of -0.046 (post-neo to post-op, N = 42) and -0.124 from post-op to follow-up (N = 33) (Wilcoxon’s signed rank test p-values = 0.036 and 0.0082, respectively) (Figure S3).

Fig. 5. Evaluation of MRD in early-stage breast cancer using MammaTrace.

A MammaTrace scores for each sample in the EBC cohort, grouped by disease status. B MammaTrace scores across treatment timepoints. (C) ROC curve showing the performance of MammaTrace for evaluating MRD at the follow-up timepoint. AUC and 95% confidence interval are indicated. D Boxplot showing the distribution of MammaTrace scores at the follow-up timepoint for DFS, REC-D, and REC-LR patients. The red dotted line represents the threshold for a MammaTrace-positive result, as determined by Youden’s index.

For the REC-D group (patients who developed distant recurrence), we observed a similar spike in the MammaTrace score between the baseline and the post-neo timepoints, with the median score increasing from 0.83 to 0.99 (N = 11, 16 samples, respectively) (Fig. 5A,B). This increase was also observed in 7 paired REC-D samples, with a median ΔMammaTrace score of 0.191 (Wilcoxon’s signed rank test p-value = 0.031) (Figure S3). However, unlike the DFS group, the MammaTrace scores in the REC-D group did not decrease at the post-op and follow-up timepoints. Median scores remained elevated at 0.93 post-op and 0.966 at follow-up (9–27 months, N = 8 and 12 collections, respectively) (Fig. 5A,B). In paired REC-D samples, MammaTrace scores showed a slight decrease from post-neo to post-op (median Δ = –0.039, p = 0.031) and from post-op to follow-up (median Δ = –0.015), though the latter was not statistically significant (p = 0.19) (Figure S3). Scores in two post-recur collections remained high at 0.885 and 0.934 (data not shown).

MammaTrace score dynamics were harder to assess in the DN-IV (de novo metastasis) and REC-LR (locoregional recurrence) groups due to the small number of patients (Table S4). In the DN-IV group, one patient had a high follow-up score (0.97), consistent with active disease, while the second patient lacked a viable follow-up sample (Fig. 4, Fig. 5A,B). In the REC-LR group, scores were high at baseline (N = 3) and post-neo (N = 4; scores = 0.985, 0.975), with a gradual decline at post-op and follow-up (N = 3 each; scores = 0.939, 0.875) (Fig. 5A,B).

Next, we used ROC curve analysis to evaluate the performance of MammaTrace in distinguishing DFS from REC-D samples at each collection time point (Fig. 5C, Figure S4). Among all timepoints, the follow-up collection ( > 9 months post-surgery) yielded the highest AUC of 0.88 (Fig. 5C, Figures S4–S5, Table S5), indicating it was the most informative timepoint for assessing MRD. Accordingly, the follow-up timepoint was selected for further analysis of the relationship between MRD status and patient survival (Fig. 5C). We then established a threshold to classify MammaTrace-positive and -negative samples at follow-up using Youden’s index, which optimizes the combined sensitivity and specificity. We selected the Youden index to provide an objective and unbiased method for threshold determination. At the resulting threshold of 0.897, 10/11 REC-D samples were MammaTrace-positive, corresponding to a sensitivity of 0.91 (Fig. 5D, Fig. 6A). While the AUC remains a valid measure of the classifier’s overall performance, sensitivity and specificity at the selected threshold may vary slightly in independent datasets due to sample-specific optimization. MammaTrace detected MRD in patients with liver, lung, bone, and brain metastases, among other metastatic sites, suggesting utility across a broad range of metastatic sites (Table S6). The one false-negative patient (E100) had mediastinal and bone metastases (Table S6). Among DFS patients, a total of 33/40 were MammaTrace-negative, with a median post-surgical follow-up of 76 months, yielding a specificity of 0.83 (Fig. 5D, Fig. 6A). The one DN-IV sample was MammaTrace-positive at follow-up. In the REC-LR group, 1/3 tested positive (Fig. 5D, Fig. 6A). REC-LR samples were not used for ROC analysis or threshold determination. Based on these findings, we determined that the follow-up interval of at least 9 months post-surgery was the most informative timepoint for MRD assessment using MammaTrace, as MammaTrace scores were most strongly correlated with subsequent distant recurrence.

Fig. 6. MammaTrace-positive results precede clinical recurrence.

A Timeline of all cfDNA sample collections for each patient, colored by MammaTrace result (red = positive, blue = negative). Patients are grouped by disease status (DFS, DN-IV, REC-D, REC-LR), and molecular subtype is annotated on the left. B Samples from all REC-D and REC-LR patients with a follow-up blood collection (N = 14). Distant recurrence is marked with a red cross (N = 11), and locoregional recurrence with a blue diamond (N = 4); one patient (E003) experienced both. The lead time (in months) between a positive MammaTrace result and clinical recurrence is labeled for each patient. C Kaplan-Meier curve showing progression-free survival stratified by MammaTrace status at follow-up. Log-rank p-value and risk table are shown.

To investigate the reasons for the false-positive results, we examined the 7 DFS patients that were MammaTrace-positive at follow-up, but did not experience clinical recurrence. False positives were not associated with molecular subtype (2 HR + /HER2-, 1 HR + /HER2 + , 4 TNBC), tumor stage (N stage I = 3, N stage II = 3, N stage III = 1), race, or body mass index (Figure S6). All 7 patients were clinically disease-free at the study’s conclusion in July 2024, with a median disease-free survival of 6.04 years (range: 5.28–8.07 years).

Lead time and prognostic value of MammaTrace for recurrence prediction

To assess the potential lead time for predicting clinical recurrence using MammaTrace, we analyzed the time interval between a MammaTrace-positive result at follow-up and the date of clinically confirmed recurrence. Among the 10 MammaTrace-positive REC-D patients with follow-up collections, the median lead time was 15.0 months (457 days), defined as the interval between a positive MammaTrace result at follow-up and clinically confirmed recurrence under standard-of-care surveillance (Fig. 6B). Two patients had short lead times: one patient had a lead time of one day (E006) and a second patient (E072) had a lead time of one month (Fig. 6B). E072 was an HR + /HER2- patient who experienced recurrence approximately one year after surgical resection, suggesting the presence of pre-existing micrometastatic disease rather than a typical distant recurrence pattern for this subtype. All remaining patients had lead times ranging from 4.4 to 24.4 months, indicating that MammaTrace may provide substantial early warning of recurrence. A MammaTrace-positive result at follow-up was also significantly associated with worse prognosis (Fig. 6C) (log-rank p-value < 0.0001) and significantly outperformed pathologic complete response (Figure S7).

Taken together, these findings show that MammaTrace identifies patients with MRD during later post-treatment surveillance who are at risk for future disease recurrence and provides a substantial lead time relative to clinically confirmed recurrence under standard-of-care follow-up.

Discussion

The present study introduces MammaTrace, a tumor-naive, cfDNA methylation-based, plasma-only liquid biopsy assay designed for MRD surveillance in breast cancer patients following curative-intent treatment. Informed by cfDNA methylation markers associated with MBC, MammaTrace demonstrated strong performance (AUC = 0.88) for MRD detection in early-stage patients. MammaTrace identified patients at increased risk of distant recurrence and provided a meaningful lead time relative to clinically confirmed recurrence. For patients who developed MBC, MammaTrace was positive in 10 of 11 cases at follow-up, with a median lead time of 15 months and up to 24 months. The single false-negative case had the longest disease-free interval (4.1 years) of any recurrent patient in the study, underscoring the importance of serial MRD testing. MammaTrace also detected a positive signal in one of the three patients who developed a locoregional recurrence. Given the small number of locoregional recurrence events in this cohort (N = 3), reliable estimation of predictive performance or sensitivity for this outcome was not possible. Importantly, MammaTrace was not designed or optimized to detect isolated locoregional recurrence, but instead to identify features associated with distant recurrence risk. Reduced sensitivity for local-only recurrence is therefore consistent with the design features of MammaTrace and with prior MRD studies^6,44. In patients who remained disease-free, MammaTrace correctly classified 33/40 as MammaTrace-negative, with a median follow-up of 76 months. Seven patients were MammaTrace-positive at follow-up but remained recurrence-free as of the study’s endpoint in July 2024, and were therefore classified as false-positive. Notably, four of these false-positive patients had TNBC, which is typically associated with early recurrence. Despite a median clinical follow-up of 76 months, delayed relapse in these TNBC false-positive patients appears unlikely, although it cannot be entirely excluded. These false-positive results may reflect biological variability in cfDNA methylation patterns, or the presence of MRD that does not progress to clinically detectable relapse within the available follow-up period. Interpretation of these results is limited by the relatively small number of recurrence events in the cohort. Further evaluation in larger cohorts with extended follow-up will be critical to better define the clinical significance of MammaTrace-positive results in disease-free patients.

To our knowledge, MammaTrace represents the first tumor-naive, cfDNA methylation-based, plasma-only MRD assay developed specifically for breast cancer surveillance following curative-intent treatment. In contrast to tumor-informed MRD approaches, which rely on sequencing of primary tumor tissue, tumor-naive assays like MammaTrace offer several advantages. Tumor-informed assays often require high-quality tumor tissue, which may not always be available or sufficient, and the process of panel development can take 6–8 or more weeks — delaying MRD detection and limiting the window for therapeutic intervention. While tumor-informed assays are highly specific, they may miss emerging or spatially heterogeneous disease due to tumor evolution or limited sampling. Tumor-naive approaches circumvent these limitations, enabling faster implementation and broader applicability, particularly in real-world settings where tissue may be scarce⁴⁵. Our findings contribute to the growing evidence base for cfDNA-based MRD detection, offering a complementary or alternative approach to existing tumor-informed assays such as Signatera (Natera Inc.) and NeXT (Personalis Inc.). Unlike tumor-informed methods that rely on tissue biopsies, MammaTrace uses a plasma-only tumor-naive approach designed to overcome these limitations. The strategy, however, is partially informed by leveraging data from cfDNA methylation patterns linked to metastasis across different breast cancer subtypes and racial backgrounds, such as from AA patients with TNBC. Conducting feature discovery directly in plasma, rather than tumor tissue — which is often unavailable, heterogeneous, or derived from different organ sites — offers a major clinical advantage. This approach makes cfDNA methylation profiling more practical and scalable for real-world MRD implementation, particularly in metastatic settings. cfDNA reflects both tumor-derived DNA and systemic factors, capturing a broader, clinically relevant signal that may be absent in tumor tissue. This enhances the assay’s ability to detect residual disease across various stages of cancer progression by identifying methylation changes that are commonly linked to the metastatic potential of breast cancer cells, without the need for prior tumor-specific profiling and offers a less complex and more broadly applicable solution for MRD detection.

Pooling cfDNA for WGBS during feature discovery was selected for cost and sequencing efficiency, enabling deep coverage across MBC subtypes while optimizing limited cfDNA input amounts. Pool sizes and sequencing depth were probabilistically calculated to ensure representation of individual-level methylation signals within the pooled data. Importantly, pooling was restricted to the feature discovery stage; identified DMRs were subsequently evaluated at the individual-sample level using targeted hybridization capture, demonstrating that selected methylation features were detectable and generalizable across patients.

The Signatera MRD assay (Natera Inc.)^{6–8,10,43,46,47} has been evaluated in EBC studies^6,36. In the initial landmark study, Coombes et al. found that Signatera was able to detect breast cancer MRD with 89% sensitivity and 100% specificity with a median lead time of 8.9 months⁶. A second EBC Signatera study showed similar results in a larger EBC cohort (sensitivity = 88%, specificity = 95%, median lead time = 10.5 months)³⁶. However, differences in study design and lead time definitions limit direct comparison. In the TRACERx study, Abbosh et al. found that the NeXT assay (Personalis Inc.) predicted lung cancer recurrence with 49% sensitivity and 96% specificity at a 4-month landmark timepoint⁵. While these early findings are promising, both MRD assays are tumor-informed and require primary tumor samples with high tumor purity to maintain sensitivity⁴⁸, as well as the creation of bespoke multiplex panels which add logistical complexity and cost.

Moreover, tumor-naive plasma-only approaches are not only lower in cost, complexity, and turnaround time^8,49, but offer a strategy for MRD detection in situations where sufficient tumor tissue is not available, which is often the case after treatment⁵⁰. Recent work by Parikh et al. found that a tumor-naive assay using plasma-only circulating DNA methylation and genomic alterations as markers (Reveal, Guardant Inc.) demonstrated comparable sensitivity and specificity to tumor-informed approaches in colorectal cancer patients^8,46. This assay was also used to predict breast cancer recurrence in early-stage patients and was found to have 80% sensitivity and 100% specificity in predicting distant recurrence (N = 5)⁴⁴.

The longitudinal design of our study enabled us to assess the prognostic value of MammaTrace at multiple timepoints throughout treatment, and to define the follow-up timepoint for analysis based on performance. Interestingly, after neoadjuvant treatment (post-neo) and surgical resection (post-op), we observed an increase in the MammaTrace score in some patients. This finding contrasts with those from somatic variant-based, tumor-informed ctDNA (mutation-based) MRD assays, which typically show clearance of ctDNA mutations during and after chemotherapy^38–43. This observation raises the possibility that methylation-based assays may capture biological processes not reflected in mutation-based ctDNA assays. The reason for the increase in MammaTrace scores following neoadjuvant chemotherapy is not fully understood but may reflect treatment-related effects, including chemotherapy-induced cell death or systemic treatment-related responses, such as immune activation or stromal remodeling. While clearance of ctDNA mutations typically reflects a reduction in tumor burden, elevated MammaTrace scores may instead capture biological processes including cellular stress responses, host tissue remodeling, or the emergence of resistant cell populations^51,52. These treatment-resistant cells may undergo epigenetic alterations that facilitate adaptation to chemotherapy. Accordingly, an increase in MammaTrace score in post-treatment plasma should be interpreted cautiously, as it may represent dynamic treatment-related methylation changes that accompany tumor cytoreduction as well as the development of resistant clones.

Importantly, patients whose disease recurred had sustained high MammaTrace scores compared with disease-free patients who became MRD negative by the defined follow-up timepoint. This suggests that clearance of altered, methylated DNA in response to chemotherapy, potentially arising from affected non-tumor tissues in addition to the cancer, may occur, whereas tumor-specific methylated DNA may persist in patients with MRD.

From a clinical perspective, elevated MammaTrace scores in the immediate post-operative period should be interpreted with caution, as they may reflect treatment-related epigenetic changes, whereas persistent MammaTrace positivity at follow-up appears to be more informative for identifying patients with MRD at increased risk of recurrence.

An important aspect of this study is the inclusion of cfDNA methylation features across all major breast cancer subtypes in the development of the MammaTrace assay. Our analysis revealed that methylation profiles varied significantly among the different molecular subtypes, consistent with prior research^53–56. Notably, this is the first study to identify subtype-specific methylation alterations detectable in cfDNA and to integrate these DMRs into a cfDNA-based methylation MRD assay that is applicable across breast cancer molecular subtypes.

Few MRD studies in solid tumors specifically address racial differences or include patients with diverse racial backgrounds in their machine learning model development. Most published MRD studies have focused on predominantly white populations, potentially overlooking genetic and epigenetic variations that could influence assay sensitivity and specificity across different racial groups. MammaTrace explicitly incorporated methylation features of cfDNA from AA patients with TNBC as well as normal cfDNA from AA healthy controls, in addition to including normal cfDNA samples from other racial and ethnic groups. Interestingly, our analysis revealed that cfDNA methylation patterns differ between AA and non-AA TNBC patients, highlighting the importance of incorporating population diversity during model development.

Consistent with this design, classifier performance during model construction and targeted validation was robust and consistent between AA and non-AA TNBC patients. During cross-validation, classifier performance also remained strong across breast cancer molecular subtypes and racial groups.

This approach enhances the inclusivity of MRD surveillance and addresses potential variations in cancer outcomes related to genetic and epigenetic differences across racial groups that need to be considered. However, given the limited number of recurrence events in the longitudinal cohort, definitive conclusions regarding MRD detection performance by race or molecular subtype cannot be drawn from the present study alone. Larger, multiethnic cohorts will be required to confirm subtype-specific and race-specific MRD performance.

The study has some limitations that warrant consideration. The relatively short median follow-up duration (12.4 months), cohort heterogeneity, and limited number of post-surgery timepoints should be considered when interpreting MRD dynamics and lead-time estimates. These factors also reflect the evolving nature of MRD surveillance in solid tumors, where standardized sampling intervals and uniform study designs have not yet been established. Moreover, while the overall cohort was sufficient to demonstrate statistical significance and robust classifier performance, sample sizes within certain subgroups were relatively small, and some cfDNA samples were not evaluable due to limited material—reducing the number of patients with complete longitudinal timepoints. This may limit broader generalizability and underscores the need for further validation in larger cohorts. In addition, the study included only a single follow-up collection per patient, which limited our ability to assess dynamic changes in MRD status over the course of long-term survivorship—such as whether MammaTrace-negative patients later become positive. Finally, although MammaTrace was trained using data from patients of diverse racial backgrounds, additional validation in larger, multiethnic cohorts is necessary to confirm performance across populations and explore any potential race-specific biases.

In conclusion, the MammaTrace assay shows substantial promise as a non-invasive, tumor-naive plasma-only method for MRD surveillance in breast cancer. Its ability to predict recurrence with a median lead time of 15 months highlights its potential to enable early intervention strategies. Future studies should aim to expand cohort diversity and further evaluate clinical utility in guiding treatment decisions.

Methods

Sample accrual

This study analyzed a total of 579 plasma samples from 331 patients, which were assigned to distinct analytical cohorts, including MBC/Healthy cohorts used for pooled whole-genome bisulfite sequencing (WGBS) feature discovery, model construction, and targeted validation, as well as a separate early-stage breast cancer (EBC) longitudinal cohort used for MRD evaluation (Fig. 1, Table S7). About 1-4 mL of blood was collected in EDTA (purple-top) tubes, and plasma was harvested and stored at -80 °C until further use. Samples from patients with breast cancer and from healthy individuals were obtained from multiple sources, including the University of Southern California and commercial vendors. The EBC longitudinal cohort was collected by Dr. Barb Pockaj at the Mayo Clinic, Arizona between May 2015 and October 2019 (Table S7). Both the MBC/Healthy and EBC cohorts comprised Asian, African American (AA), Hispanic/Latino (HL), and non-Hispanic White (NHW) patients (Figure S1). A summary of patient demographics for both cohorts can be found in Figure S1, Table S4, and Table S8. The MBC/Healthy cohort included plasma from MBC patients (N = 95), and healthy female donors (N = 110) (Fig. 1, Figure S1, and Table S8). In addition, we included 21 samples from African American patients with triple-negative breast cancer (AA TNBC) across varying stages during the feature discovery phase (Fig. 1, Table S9). These samples were incorporated to improve representation of diverse populations and to identify AA TNBC-associated differentially methylated regions (DMRs) independent of metastatic status. The AA TNBC pool was compared to a healthy AA cfDNA pool (N = 20 individuals) to control for race-specific methylation differences and was not compared to metastatic samples. The median age of MBC patients was 52.5 years (Figure S1), the median age of AA TNBC patients was 53 years, and the median age of healthy donors was 52 years (Figure S1).

For the EBC cohort, blood samples were collected from 107 patients (355 total samples) at 4 primary timepoints: at diagnosis (baseline), after neoadjuvant chemotherapy (post-neo), within 9 months post-surgery (post-op), and during a follow-up visit at least 9 months post-surgery (follow-up) (Fig. 4A, Table S4/S7). Blood was also collected from two patients after distant recurrence (‘post-recur’, Fig. 4A). Seventy-six samples were not used due to insufficient cfDNA yields, leaving 279 samples for analysis (Table S10). The stage breakdown is as follows: stage I N = 18, stage II N = 61, stage III N = 26 (Figure S1). Additionally, 2 patients presented with de novo stage IV breast cancer (DN-IV). All major molecular subtypes of breast cancer were represented: HR + /HER2- (N = 38), HR-/HER2+ (N = 8), HR + /HER2+ (N = 22), and TNBC (N = 39) (Figure S1, Table S4). The EBC cohort comprised 84 NHW, 12 HL, 8 Asian, and 1 AA participants (Figure S1, Table S4). Clinical outcomes for all the patients were tracked through July of 2024. Over the duration of the study, 18 patients had a distant recurrence (REC-D), 4 had a local or regional recurrence (REC-LR), and 83 patients were classified as disease-free survivors (DFS) as of July of 2024 (Figure S1, Table S4). The median disease-free survival time was 6.03 years from the end of primary treatment (min = 0.167 years, max = 8.38 years); 59 DFS patients were disease-free for more than 5 years, 24 were disease-free for less than 5 years (Figure S1). A longitudinal blood collection timeline plot is shown in Fig. 4B and a Venn Diagram showing the overlap of patients across timepoints is shown in Fig. 4C. Excluding the 2 post-recur collections, 31 patients had blood collected at 1 time point only, 18 had blood collected at 2 time points, 26 had blood collected at 3 time points, and 32 had blood collected at 4 time points (Fig. 4C).

This study was conducted in accordance with the Declaration of Helsinki and approved by the University of Southern California (USC) Institutional Review Board (IRB protocol number HS-17-00176). All patients provided written informed consent.

Whole genome bisulfite sequencing and differential methylation analysis

To investigate cfDNA methylation changes in MBC, we obtained 41 MBC samples and generated subtype-specific cfDNA pools (Fig. 1, Table S1). DNA was extracted using the MagMAX Cell-Free DNA Isolation Kit (Applied Biosystems). For WGBS, subtype-specific pools were created by combining equal amounts of cfDNA per sample. Bisulfite conversion was performed using the EpiTect Fast DNA Bisulfite kit (Qiagen: cat#59824). Sequencing libraries were prepared using the Ovation Ultralow Methyl-Seq Library kit (Nugen: cat#0335). These MBC pools included five subtype-specific groups: an HR + /HER2- pool (N = 13 individuals), an HR-/HER2+ pool (N = 8 individuals), an HR + /HER2+ pool (N = 9 individuals), and two triple-negative breast cancer (TNBC) pools (N = 5, 6) (Fig. 1, Table S1). We also generated a pool of 21 samples from AA patients with TNBC (AA TNBC) across varying stages (Table S9). Additionally, three control pools were prepared using cfDNA from healthy individuals, consisting of two predominantly NHW and HL pools (N = 9 and 10 individuals) and one AA pool (N = 20 individuals) (Fig. 1, Table S1).

Next, we performed unpooled WGBS on 21 baseline cfDNA samples from the EBC cohort (Fig. 1), comprising patients who developed distant recurrence (REC-D, N = 10), locoregional recurrence (REC-LR, N = 2), and those who remained disease-free for more than 5 years (DFS, N = 9) (Table S2). Samples from all other timepoints for these patients were not included in model training. Each cfDNA sample underwent WGBS as previously described²⁸ and was sequenced to a median depth of 22X for individual EBC samples and 53X for the MBC/Healthy cfDNA pools.

Metilene and DSS were used to identify DMRs by comparing each MBC subtype pool with healthy control groups, and by comparing EBC baseline REC-D and REC-LR samples to the DFS and healthy cfDNA pools^57,58. The AA TNBC pool was specifically compared with an AA healthy pool consisting of cfDNA from 20 individuals. Statistically significant DMRs were filtered using defined thresholds for methylation difference, adjusted p-values, and minimum coverage depth to ensure robust and reproducible identification. In addition to DMR analysis, we profiled cfDNA fragment-level methylation patterns. We defined CpG groups as sets of 2–5 consecutive CpGs separated by ≤100 bp across the genome. For each group, we tabulated the methylation states of all cfDNA reads spanning the locus in MBC, EBC, and healthy samples. We then selected 1300 groups (800 from MBC, 500 from EBC) whose fragment-level methylation patterns were enriched in MBC relative to EBC and depleted in healthy cfDNA.

To generate the final feature set, we applied filters for |Δβ | , minimum group-wise coverage, and statistical significance (p/q-value thresholds). We then integrated all DMR lists, harmonized coordinates to autosomes and bounded width (chr1–22, ≤1 kbp), and collapsed overlaps into a single nonredundant union. This produced one consolidated panel comprising 5077 regions designed to preserve biologically relevant contrasts across cohorts, subtypes, and analytical approaches.

Hybrid probe capture

To investigate the potential for the 5077 MBC-associated DMRs as markers of MRD for EBC, we designed a custom hybridization probe capture assay, to selectively enrich each of those regions. In total, the 5077 DMRs spanned approximately 1.24 Mb and encompassed 71,118 CpG loci, which were used to design a targeted hybridization probe capture panel (Daicel Arbor Biosciences). WGBS read data from MBC and EBC WGBS samples, exhibiting a range of methylation states, were used as the reference sequence for probe design. Sequencing reads aligning to the target regions were extracted and used to guide the design of an 80-nucleotide probe set with approximately 3X tiling density. Redundant or highly similar probe sequences were filtered out to reduce overlap and ensure efficient target enrichment. The final capture panel comprised approximately 474,000 probes.

Hybridization probe capture, sequencing, alignment, and methylation calling on all samples were performed as previously described^34,35,59. Six nanograms of cell-free DNA was concentrated to a volume of 13 microliters and used as template for bisulfite conversion and library preparation using the Tecan Ultralow Methyl-Seq kit following the manufacturer’s protocol (Tecan Genomics: cat#9513-A01). Final library products were quantitated using TapeStation D1000 High Sensitivity screen tapes (Agilent: cat#5067-5584). Eight libraries were combined for targeted capture using custom myBaits by Daicel Arbor Biosciences according to manufacturer protocol. The amplification product of the first hybridization probe capture was used as template for a second round of capture/wash/amplification. Final amplification products were quantitated using TapeStation D1000 High Sensitivity screen tapes. Equimolar pools of capture products were sequenced on the NovaSeq6000 (Illumina) at the Keck Genomics Platform. For NGS quality control (QC), alignment rate was calculated by Bismark⁶⁰, and cfDNA insert size distributions were calculated by qualimap bamQC⁶¹; all QC metrics were aggregated using MultiQC⁶².

Machine learning

A machine learning classifier was trained to distinguish MBC samples from healthy controls for subsequent MRD detection. A support vector machine (SVM) binary classifier was trained using sequencing data from patients with MBC (N = 72) and healthy individuals (N = 83) (Fig. 1, Table S3). A subset of these samples (13 MBC and 12 healthy individuals) had previously contributed cfDNA to pooled WGBS feature discovery. However, WGBS data were used exclusively to identify candidate methylation features, whereas model training and targeted validation were conducted using independent targeted-capture sequencing data. This separation ensured that model training was not influenced by the feature discovery dataset. Five additional non-MBC AA TNBC samples were also included during model construction and targeted validation to enhance subtype and racial representation; these samples were not labeled or treated as metastatic cases during model training.

To enrich for ctDNA, cfDNA sequencing data were preprocessed to retain only read pairs with fragment lengths below 150 bp, a threshold previously associated with ctDNA enrichment and improved classifier performance^63,64. Next, for each target region, we computed the proportion of unmethylated (U), mixed (X), and methylated (M) reads mapped to each region. This fragment-level “UXM” approach has demonstrated utility in cfDNA-based tissue deconvolution⁶⁵. Of the 5077 identified DMRs, 4049 regions passed quality control filtering and were used as features for model training. Model training and evaluation were conducted using five repetitions of five-fold cross-validation to ensure robust performance assessment⁶⁶. After cross-validation, a final model — referred to as the MammaTrace model — was trained using the full dataset and subsequently applied to the EBC cohort for MRD detection. The output of this model, referred to as the MammaTrace score, is a continuous value ranging from 0 to 1, with higher values reflecting a cfDNA methylation profile characteristic of MBC and associated with an increased risk of distant recurrence. All analyses were performed in R version 4.3.2 and summarized/visualized using the dplyr, tidyr, ggplot2, ggpubr, caret, pROC, yardstick, SummarizedExperiment, and ComplexHeatmap packages^67–72.

Author contributions

Study conception: B.S. experimental design: B.S., B.P., D.B., G.G. data generation: B.S., D.B., A.K., G.G., M.P., J.G., B.P. data analysis: B.S., D.B., J.P.L. manuscript preparation: D.B., A.K., G.G., J.P.L., D.Sp., J.C., D.St., H.J.L., C.H.H., C.L., J.O., B.P., B.S. All authors read and approved the final manuscript.

Data availability

Methylation sequencing data have been deposited on Zenodo (https://doi.org/10.5281/zenodo.21349212).

Code availability

The code generated in this study is not publicly available. Access may be granted upon reasonable request at the discretion of the investigators and dependent on the proposed use.

Competing interests

B. Salhia is founder and shareholder of CpG Diagnostics and reports other support from CpG Diagnostics during the conduct of the study, as well as personal fees from AstraZeneca and Illumina outside the submitted work; in addition, B. Salhia and D. Buckley have a patent “Cell-Free DNA Methylation Test for breast cancer # PCT/US2023/081012” pending and licensed to CpG Diagnostics Inc and may receive royalty payments. G. Gooden is an advisor to CpG Diagnostics. USC has an ownership interest in CpG Diagnostics and the potential to receive royalty payments from CpG Diagnostics. All other authors declare no competing financial or non-financial competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41698-026-01595-8.