Clinical Validation of Digital PCR-Based ctDNA Detection for Risk Stratification in Residual Triple-Negative Breast Cancer: TRICIA Trial Results.

Introduction
Triple-negative breast cancer (TNBC) is the most aggressive form of breast cancer and accounts for about 15% to 20% of all breast cancers (1). The term “triple negative” refers to the lack of protein expression of the three most common biomarkers in the treatment of breast cancer: estrogen receptors, progesterone receptors, and HER2. The lack of these receptors makes TNBC unresponsive to common hormonal therapies or HER2-targeted treatments, leaving chemotherapy as the mainstay of treatment. In early-stage TNBC, neoadjuvant chemotherapy (NAC) is used to reduce the size of the tumor before surgery. The extent of response to NAC is a strong prognostic marker; patients with residual tumor at the time of surgery [nonpathologic complete response or (non-pCR)] have a significantly higher risk of relapse and death. About half of patients with TNBC treated with NAC have residual disease, and 35% of these patients show disease recurrence within 2 years (2). The residual cancer burden (RCB) score, which measures the extent of residual disease in the tumor and lymph nodes, serves as an additional prognostic tool to stratify recurrence risk (3). The CREATE-X clinical trial (4) demonstrated that the addition of adjuvant chemotherapy (capecitabine) results in improved survival in patients with non-pCR. However, only about 15% of patients derive benefit from this additional chemotherapy (4), suggesting that the majority, around 85% of these patients, may be exposed to unnecessary toxicity with no improvement in outcomes. Identifying patients who can truly benefit from further adjuvant chemotherapy is an unmet need in the care of patients with TNBC.
Recently, plasma-based biomarkers or liquid biopsies have emerged as possible cancer biomarkers (5). One such biomarker is circulating tumor DNA (ctDNA). ctDNA is short tumor-derived fragments of DNA that circulate freely in serum and plasma (6) and represents a small proportion of total cell-free DNA (cfDNA; ref. 7). Measuring ctDNA from serial blood samples can provide real-time information about the disease and treatment progression (8). The measurement of ctDNA is ultra-specific (detecting DNA variants present only in tumor cells) and very sensitive [with levels of detection of ctDNA in the range of 0.1%–10% of total circulating free DNA (9)]. Previous work has shown that the detection of ctDNA after surgical resection of the primary breast tumor indicates poor prognosis, suggesting the presence of minimal residual disease (MRD) in the body (10). The detection of minute quantities of ctDNA indicative of MRD requires tumor-bespoke or tumor-informed assays, in which personalized assays are developed based on DNA sequencing of each tumor. Using such a tumor-bespoke assay developed in an academic setting, we recently showed that the presence of detectable ctDNA after chemotherapy but before surgery was a reliable indicator of poor prognosis in both patients with pCR and non-pCR TNBC (11).
The TRICIA trial (NCT04874064) focused on patients with TNBC with non-pCR, with the aim to investigate tissue- and plasma-based biomarkers associated with resistance to chemotherapy in this disease. A key part of its objectives is to both validate and extend our previous findings of the strong prognostic value of post-NAC/preoperative ctDNA testing. In addition, we included blood collected at three different postoperative time points. During the study, the standard of care changed, and patients received adjuvant capecitabine. The analysis of serial plasma samples in these patients receiving capecitabine forms the largest cohort of ctDNA taken before, during, and after adjuvant capecitabine treatment in patients with TNBC. In the current study, we validated the very strong prognostic value of the post-NAC/preoperative time point (T1) and demonstrated the prognostic value of ctDNA clearance during capecitabine treatment.

Materials and Methods

Patient recruitment
Female patients (individuals assigned female at birth) were invited to participate in the TRIple Negative Breast Cancer markers in Liquid Biopsies using Artificial Intelligence trial (TRICIA trial; NCT04874064) if they had TNBC and had residual tumor after standard-of-care NAC. The representativeness of the study participants can be found in Supplementary Table S1. Post-NAC/preoperative blood draws were done routinely prospectively on all patients receiving NAC at the Jewish General Hospital (JGH). At other sites, radiologic response was used to preidentify non-pCR patients. Sixty-nine patients were recruited in three different centers in Canada, namely the JGH, Ottawa Hospital Research Institute, and London Health Sciences Center between 2020 and 2022. The last participant was enrolled in August 2022. This study was approved by the Research Ethics Board (REB) from the JGH (protocol MP-05-2020-1813) and the Ontario Cancer REB (CTO Project ID: 1934).
To expand our TRICIA cohort, we also included 48 patients with retrospectively collected samples and the same inclusion criteria from the JGH breast biobank (JGH REB protocol 05-006). In accordance with the Declaration of Helsinki, all patients provided written informed consent. A total of 117 patients were consented and 92 were included in this study (CONSORT Diagram Supplementary Fig. S1). For one of the patients, there were two different TNBC tumors (one from each breast) that were sequenced and ctDNA assays developed, and the results were included as independent ctDNA patient results for all statistical analyses. This is a real-world study in which all patients were high risk (non-pCR) and were followed by the treating oncologists in each center as per standard of care. Treatment decisions were made independent of ctDNA findings. About one-third of the way into our study (May 2016), adjuvant capecitabine became the standard of care for patients with non-pCR TNBC at the JGH and London Health Center. Surveillance intervals were completely independent of ctDNA status as the clinicians were not made aware of their patient’s results. ctDNA analysis was batched following the end of treatment for each case. Patient clinical outcomes [relapse-free survival (RFS) and overall survival (OS)] were concealed until the time of analysis.

Blood collection and processing
Serial blood samples were collected at different time points: after NAC but before surgery (T1), early postoperatively (T2), and late postoperatively or after adjuvant treatment (T4). For 67 patients receiving capecitabine, T2 blood samples were taken prior to capecitabine treatment, a T3 collection was added during capecitabine, and T4 blood samples for these patients were always collected after completion of adjuvant capecitabine treatment. For each time point, blood was collected in 6-mL k-EDTA collection tubes and processed within 2 hours of collection following standard operating procedures. The samples were centrifuged at 2,500 rpm for 15 minutes at room temperature. Plasma and buffy coats were aliquoted into barcoded 1.5-mL tubes and stored at −80°C. Before cfDNA extraction, aliquots were thawed and subjected to a second centrifugation at 10,000 rpm at room temperature.
Commercial pooled human plasma (blood derived, IPLAWBK2E1000ML) was purchased from Innovative Research (lot #36336), aliquoted, and frozen at −80°C. cfDNA was extracted and used as normal plasma control to measure the background fluorescence and to establish variant allele frequency (VAF) detection thresholds for all tested assays.
cfDNA extraction was performed from 2 mL of plasma using our modified hybrid JGH extraction protocol as previously reported (11, 12). DNA was extracted from buffy coats using the QIAamp DNA Mini Kit (Qiagen, cat. #51304).

Tumor whole-exome sequencing and variant calling
Formalin-fixed, paraffin-embedded (FFPE) samples of residual tumors or diagnostic biopsies were used for whole-exome sequencing (WES). For DNA extraction using the QIAamp DNA FFPE Advanced UNG Kit (Qiagen, cat. #56704), 10-µm sections were used. Macrodissection was performed to enrich for tumor cells when required. All extractions were performed on tissue containing at least 50% tumor cellularity confirmed by a pathologist (MP).
For library preparation, 200 ng of gDNA from tumor and lymphocytes was used at the Institute for Research in Immunology and Cancer’s Genomics Platform. Libraries were prepared using the SureSelect XT with the Enzymatic Fragmentation Library Prep Kit (Agilent) according to the manufacturer’s instructions. Exome capture was performed using the SureSelect Human All Exon v7 Library Prep Kit (Agilent). Libraries were quantified by Qubit and Bioanalyzer DNA 1000. All libraries were diluted to 10 nmol/L and normalized by qPCR using the KAPA Library Quantification Kit (KAPA, cat. #KK4973). Libraries were pooled to equimolar concentration. Sequencing was performed at the McGill Genome Center on the Illumina NovaSeq S4 200-cycle flowcell (2 × 100 bp).
Preprocessing, including removal of the molecular barcode, adapter sequences, and low-quality sequences, was performed using fastp. Reads were then aligned to the hg37 genome using the Burrows–Wheeler Aligner BWA-MEM tool. The variants were called using the GATK4 best practices workflow for somatic short variant discovery (https://gatk.broadinstitute.org/hc/en-us/articles/360035535912-Data-pre-processing-for-variant-discovery). Additionally, a panel of normal sequences using all lymphocyte germline data from the study was created to capture technical artifacts. The GATK4 best practices workflow was followed up to the FilterMutectCalls step (https://gatk.broadinstitute.org/hc/en-us/articles/360035894731-Somatic-short-variant-discovery-SNVs-Indels). Finally, variants were annotated using snpEff tool (http://pcingola.github.io/SnpEff/). Mutect2 VCF output includes a two log-OR confidence score indicating the likelihood that the tumor is present. BCFtools was used to filter for only variants that passed the GATK4 tool scoring (Tumor Logarithm of the Odds or TLOD > 6.3). The data were then passed through the following filters: a mapping quality above 40, a read depth of VAF above 10 reads, and the variant being absent from the panel of normal sequences. The top five somatic variants per patient were selected for digital droplet PCR (ddPCR) assay development based on the highest TLOD score and the highest VAF. Only variants with VAF greater than the median VAF values of all filtered variants were considered for each patient.

Development of ddPCR assays and measurements
Five ddPCR assays were developed per patient using a set of compatible primers and probes for each variant designed and ordered from Integrated DNA Technologies (see Supplementary Table S2). ddPCR conditions were optimized to determine the optimal annealing temperature for each assay, ranging from 52°C to 62°C, using the C1000 Touch Thermal Cycler with a 96–deep well reaction module (cat. #80224, Bio-Rad). Each ddPCR assay was validated using DNA extracted from the respective tumor.
According to the manufacturer’s instructions, 20 ng of cfDNA was used for targeted preamplification reactions with the SsoAdvanced PreAmp Supermix (cat. #1725160, Bio-Rad). Detection of mutant variants was performed using the Droplet Digital PCR system (QX200 Droplet Generator, cat. #1864002; PX1 PCR Plate Sealer, cat. #1814000; C1000 Touch Thermal Cycler with 96–deep well reaction module, cat. #80224; and QX200 Droplet Reader, cat. #17005228; Bio-Rad) and ddPCR Supermix for probes (no UTP; cat. #Q33216, Bio-Rad). The final concentration of the primer/probe mix was 900 nmol/L/250 nmol/L. Preamplified DNA was diluted 100-fold to achieve optimal resolution. For ddPCR reactions, 8 µL of diluted preamplified DNA per well was used. The thermal cycling protocol was 10 minutes at 95°C (30 seconds at 94°C and 1 minute at Ta °C) × 50 cycles; 10 minutes at 95°C; and hold at 4°C. QX Manager 2.2 Standard Edition was used for cluster determination and VAF calculation.
All samples were run in triplicate and each run included normal plasma control cfDNA, the patient’s lymphocyte DNA to control for clonal hematopoiesis of indeterminate potential, and tumor DNA as a positive control. For some patients, additional positive controls included pre-NAC treatment cfDNA samples and cfDNA samples from the metastatic stage. All controls were run in triplicate.
A variant was considered detectable if the VAF (proportion of mutant alleles/nonmutant alleles) was greater than at least twice the SD of the VAF of the normal plasma controls (11) and if there were at least three positive droplets across replicates.
Corrected fractional abundance (cFA) was calculated by subtracting the threshold of detection (variant-specific threshold of detection) from the average fractional abundance (FA) of a given time point sample of each variant (in triplicate).

Analytic validation of the mutation detection protocol
We evaluated the analytic performance of our assays using a representative test panel. This panel included 16 assays, including five used in this TRICIA project and 11 newly designed from commercial reference material (Supplementary Table S3). Of note, as with the entire TRICIA cohort, we used only 2 mL of plasma for these tests.
The within-run precision coefficients of determination calculated for the linear regressions performed between two random replicates of the three were R2 = 0.996 (n = 101; AF, 0.005%–5%; Supplementary Fig. S2A). Overall, within-run precision of the panel of 16 assays showed SD = 0.052 and coefficient of variation (CV) = 17.8% (Supplementary Table S4). The between-run precision coefficients of determination calculated for the linear regressions performed between an average of triplicate mutant AF were R2 = 0.929 (n = 93; AF, 0.005%–5%; Supplementary Fig. S2B). Overall, between-run precision of the panel of 16 tested assays showed SDE = 0.04 and CV = 16.2% (n = 63; AF, 0.005%–5%; Supplementary Table S5).
For sensitivity and specificity assessment, we tested the panel of reference materials with a range of VAF from 0.005% to 5% and wild-type samples only, as well as cfDNA from two healthy donor plasmas. Results of sensitivity calculation are shown in Supplementary Table S6 and Supplementary Fig. S2C. As expected, higher VAF means a higher chance of detecting the presence of mutant alleles. At the tested range of VAF 0.005% to 5%, overall variant detection shows sensitivity = 80% and specificity = 99%. Specifically, the range of 0.1% and above, which covers 69% of detectable time points of the TRICIA study, demonstrated excellent sensitivity (97%) and specificity (99%). The lowest VAF range (0.005%–0.05%) shows a notable drop in sensitivity (41%) while maintaining high specificity (99%; Supplementary Table S6).

RCB score calculations
RCB scores were evaluated using the criteria described by the MD Anderson Cancer Center (https://www.mdanderson.org/content/dam/mdanderson/documents/for-physicians/clinical-calculators/PLM_protocol_RCB_calculators_BRO.pdf).

Statistical analysis
Kaplan–Meier survival curve analysis was performed for RFS, which was calculated from the time of surgery to the time of relapse or last follow-up. OS was calculated from the time of diagnosis to the time of death or last follow-up. Kaplan–Meier curves were generated for RFS and OS using GraphPad Prism (RRID: SCR_002798).
Kaplan–Meier curves were analyzed for statistical differences between ctDNA status, capecitabine treatment, and RCB score analyses using log-rank (Mantel–Cox) tests for significance and log-rank tests for calculating hazard ratio (HR) with confidence intervals (CI) in GraphPad Prism (RRID: SCR_002798).
Bar graphs and violin plots were analyzed by unpaired two-tailed t tests for comparing cFA between two groups (relapse vs. nonrelapse, ctDNA positive vs. negative, RCB scores, and time points) or one-way ANOVA for comparing cFA between three or more groups (RCB scores and time points). Paired two-tailed t tests were used for comparing changes in cFA values of matched samples. To evaluate the correlation between tumor cFA and tumor VAF, the parametric Pearson correlation (r) test was used. The Fisher exact test was used for comparing proportions (stacked bar graphs associated with Kaplan–Meier plots). All statistical analyses were performed in GraphPad Prism (RRID: SCR_002798). The multivariate modeling and proportional hazard assumption were calculated using R project software (RRID: SCR_001905).
Performance metrics were calculated as follows:Positive predictive value = ctDNA+ relapses/(ctDNA+ relapses + ctDNA+ nonrelapses)

Negative predictive value = ctDNA− nonrelapses/(ctDNA− nonrelapses + ctDNA− relapses)

Sensitivity = ctDNA+ relapses/(ctDNA+ relapses + ctDNA− relapses)

Specificity = ctDNA− relapses/(ctDNA− relapses + ctDNA+ nonrelapses)

Local relapse is defined as a recurrence in the breast tissue or axillary lymph nodes on the same side as the primary tumor. Distant relapse is a recurrence at a distant site away from the original breast tumor. Lead time was calculated from the earliest postoperative ctDNA detection (T2, T3, or T4).
This study followed the REMARK biomarker reporting guidelines.

Results

Patient characteristics
A total of 117 patients with TNBC with non-pCR following NAC were part of the TRICIA study cohort. Ninety-two patients with high-quality tumor WES data and blood collected at least one of the four prespecified time points were included in this ctDNA study (Supplementary Fig. S1). Blood was collected at the following time points: T1 after NAC but before surgery (median, 0.43 months before surgery; range, 0–4.6 months), T2 early after surgery (median, 1.1 months; range, 0.03–4.1 months), T3 during adjuvant treatment (median, 3.8 months; range, 1.4–8 months), and T4 late postoperatively or at the end of adjuvant treatment (median, 8.7 months; range, 3.5–15 months; Fig. 1A). The T3 time point was only collected in patients who received adjuvant capecitabine.
The clinical characteristics of this cohort are found in Table 1. The median age range of the patients was 55.5 years (range, 28–79), and 92% had stage II or III disease at the time of diagnosis. Most patients received standard adriamycin/cyclophosphamide/paclitaxel NAC (87%), with carboplatin added in 41% of cases and atezolizumab in 10% of cases. The median follow-up of these patients was 38.2 months (12.2–236.4 months). At the time of data cutoff, 37 (40%) patients had relapsed (31 distant and six local relapses) and 27 patients had died. Sixty-one of 92 patients (66%) received adjuvant capecitabine and among these, 20 patients (33%) experienced relapse. Consistent with the findings of the CREATE-X trial (4), patients receiving capecitabine had a better 5-year RFS (66% vs. 41%, capecitabine vs. no capecitabine, respectively; HR, 0.47; 95% CI, 0.23–0.96; P = 0.02). OS was also better, but the difference was not statistically significant (64% vs. 48%, capecitabine vs. no capecitabine, respectively; HR, 0.51; 95% CI, 0.22–1.14; P = 0.072; Supplementary Fig. S3).
Tumor-informed personalized ctDNA assays were developed in-house with five variants per tumor for a total of 467 variants (Supplementary Table S2). Each assay was first validated in the matched patient tumor (Supplementary Fig. S4) and then used to analyze 262 plasma samples (Supplementary Table S3). We controlled for clonal hematopoiesis by using the patient’s normal control sample, and only three variants were identified as clonal hematopoiesis mutation and replaced by additional tumor-specific variants. In total, we selected 430 genes, and 92% were unique to each patient; only TP53 (n = 20) and PIK3CA (n = 4) were selected in more than two patients (Supplementary Table S2). ctDNA was defined as positive if at least one variant was detectable at each time point tested. The proportion of the number of variants detected per patient at each time point is shown in Supplementary Fig. S5. Figure 2 shows a patient summary throughout clinical follow-up, describing the ctDNA detectability at each time point and patient status.

Validation of post-NAC/preoperative ctDNA detection as a strong predictor of patient outcomes
We initially assessed the ctDNA status at the end of NAC and before surgery to provide an independent validation of the strong prognostic value of ctDNA detection using our assays at this time point, as previously reported by our group (11). We analyzed T1 samples from 64 patients. Among these, 44 patients were classified as ctDNA positive (69%), and 20 patients were classified as ctDNA negative (31%). During the follow-up period, 29 patients experienced disease relapse: three had local relapse, whereas 26 had distant relapses. The number of detectable variants per patient at T1 varied according to patient outcomes; no variants were detectable in 54% of nonrelapsed patients compared with 3% of relapsed patients, and one to two variants were detectable in 40% of nonrelapsed patients compared with 55% of relapsed patients.
The overall recurrence rate in ctDNA-positive patients was 64% (28 of 44) and 5% (one of 20) in ctDNA-negative patients at T1 (P < 0.001; Fig. 1B). This means a negative predictive value (NPV) for no relapse of 95% and a positive predictive value (PPV) for relapse of 64%. These predictive values closely align with those reported in our previous study (11), in which we observed an NPV of 89% and a PPV of 71%. A T1-positive ctDNA test was detected in 28 of 29 patients who relapsed, yielding a high sensitivity of 97%. In contrast, 19 of 35 patients who did not relapse tested negative for ctDNA, resulting in 54% specificity.
The RFS for ctDNA– patients was significantly longer than in ctDNA+ patients (median, 55.7 vs. 19.5 months, respectively; P < 0.0001), with an HR of 0.052 (95% CI, 0.024–0.109), reflecting a lower likelihood of relapse (Fig. 1B). Similarly, OS for ctDNA– patients was markedly improved (P = 0.0001), with an HR of 0.06 (95% CI, 0.027–0.131), further indicating the favorable prognosis in this cohort (Fig. 1C). In fact, the 5-year OS of the ctDNA– cohort was 95% compared with 35% for the ctDNA+ cohort. CtDNA status and clinical outcomes are summarized in Supplementary Table S7. The consistency between the findings of this independent cohort and our earlier work further validates the prognostic value of ctDNA at this post-NAC/preoperative time point as measured by our approach.
We repeated the analysis only on patients receiving adjuvant capecitabine (n = 39). As with the results for the entire cohort, the T1 time point was highly prognostic for RFS (P = 0.0009) and OS (P = 0.001; Fig. 1E and F). The sensitivity and specificity for relapse detection were 100% and 54%, respectively. A total of 38% of patients were ctDNA– and the NPV was 100% whereas the PPV was 58%. The 5-year OS for ctDNA– patients at T1 was 100%.
We further investigated the relationship between ctDNA levels and relapse looking at the FA of each variant at the T1 time point. We derived cFA values by subtracting the FA in normal pooled plasma from the FA in the patient sample for each variant (Supplementary Tables S8 and S9). Any negative cFA (i.e., values that were lower than the normal plasma pool threshold) were considered as 0 (undetectable). We then compared the mean cFA values between patients who relapsed and those who did not relapse at the T1 time point. The mean cFA for patients who relapsed was 2.1% (range, 0.0017–20.9) compared with 0.12% (range, 0.0005–0.79) in patients who did not relapse (unpaired t test P = 0.0275; Fig. 1D). Therefore, higher amounts of ctDNA reflected in higher cFA levels at the T1 time point are associated with a greater likelihood of relapse.

Post-NAC/preoperative ctDNA can distinguish poor- and good-outcome patients within RCB 2 and RCB 3 groups
One of the strongest prognostic markers in patients with non-pCR TNBC is RCB, classifying patients according to the extent of residual cancer at the time of surgery in both the breast and the lymph nodes following neoadjuvant treatment (13). Of the 92 patients, 21 (23%) patients were RCB 1 (MRD), 45 (49%) patients were RCB 2 (moderate residual disease), 25 (27%) were RCB 3 (extensive residual disease), and one (1%) patient did not have an RCB score. As expected, and as previously shown in patients with TNBC (3, 14), RCB 1 patients had a higher RFS and OS, followed by RCB 2 and RCB 3 patients (Supplementary Fig. S6B). Two (10%) RCB 1 patients, 17 (38%) RCB 2 patients, and 18 (72%) RCB 3 patients relapsed (Supplementary Fig. S6A). Of note, the RCB 1 patients had an estimated 81% 5-year RFS in this cohort, similar to what was previously reported in patients with TNBC (3, 14).
At the post-NAC/preoperative time point (T1), we observed a gradual increase in the proportion of ctDNA+ patients with increasing RCB score (Supplementary Fig. S6C). We also found that the average ctDNA levels (cFA) increased with RCB score: 0.11 (RCB 1), 0.38 (RCB 2), and 2.85 (RCB 3; P = 0.006; Supplementary Fig. S6D). These results confirm that ctDNA detection and quantity correlate with residual post-NAC tumor burden in the breast and lymph nodes after NAC and before surgery.
We next associated ctDNA detection prior to surgery with patient outcomes (RFS and OS) for each RCB category separately. Of the two RCB 1 patients who relapsed, one had a plasma sample at the T1 time point, which was ctDNA+. ctDNA detection at T1 was not significantly associated with RFS or OS (Fig. 3A and B). However, for RCB 2 patients, ctDNA detection at T1 was significantly associated with RFS [P = 0.04, HR, 0.156 (95% CI, 0.048–0.503)] and OS [P = 0.032, HR, 0.150 (95% CI, 0.045–0.503); Fig. 3C and D]. In fact, of 30 RCB 2 patients with T1 samples collected, 12 patients developed disease relapse, and 92% (11 of 12) were ctDNA+ at T1. Nine patients were ctDNA– and only one patient relapsed, whereas 11 of 21 (52%) ctDNA+ patients relapsed. In patients with an RCB 3 score, presurgical detection of ctDNA was strongly correlated with both RFS and OS (P = 0.0008 and P = 0.005, respectively; Fig. 3E and F). Notably, 100% of RCB 3 patients who were ctDNA– at T1 remained relapse free (100% specificity) and alive at the time of data cutoff, whereas every RCB 3 patient with detectable ctDNA experienced disease relapse (100% sensitivity). All together, these findings highlight that ctDNA can further stratify RCB 2 and RCB 3 patients into poor- and good-outcome groups, suggesting that ctDNA detection can add significantly to prognosis above and beyond RCB, especially in patients with moderate-to-extensive residual disease after NAC. For this reason, we pooled data from RCB 2 and RCB 3 patients together and reanalyzed the prognostic value of presurgical ctDNA. The absence of ctDNA detection at the T1 time point was associated with a 5-year RFS of 92%, whereas the detection of ctDNA was associated with a 5-year RFS of 25% (P = 0.0002, Fig. 3G). ctDNA+ patients had an extremely poor 5-year OS, with only 10% patients alive compared with 90% for ctDNA− patients (P = 0.0003; Fig. 3H). In this combined group, the test demonstrated strong performance characteristics, with a PPV of 73%, NPV of 92%, and sensitivity and specificity of 96% and 55%, respectively. These findings support the use of presurgical ctDNA as a unified prognostic test across both RCB 2 and RCB 3 patients to identify low-risk patients who may be spared further adjuvant treatment.

CtDNA dynamics and clearance during surgery
As ctDNA can reflect tumor burden changes after surgical intervention, we examined the ctDNA dynamic changes from preoperative (T1) to early postoperative (T2) samples. Among the 39 patients with plasma collected at both T1 and T2 time points prior to any adjuvant treatment, 26 were ctDNA+ at T1. Of these, eight (31%) patients became ctDNA– after surgery. On the other hand, seven patients who were initially ctDNA− at T1 became ctDNA+ at T2 (Supplementary Fig. S7A). The log-rank test showed statistically significant changes in RFS among the different ctDNA detection groups (P = 0.0207; Supplementary Fig. S7B). Patients with no ctDNA detectable at T1/T2 (−/−) had 100% RFS at our median follow-up time point of 38 months along with those patients who converted to positive at T2 (−/+) compared with 33% RFS in persistent ctDNA (+/+) patients (P = 0.026 and P = 0.017, respectively). Cox regression analysis showed that patients who cleared their ctDNA (+/−) following surgery tended to have better RFS at median follow-up (63% relapse rate) than patients who did not clear (+/+; P = 0.47) but poorer RFS than −/− and −/+ patients (P = 0.107 and P = 0.082, respectively). These data suggest that prognosis was more determined by the presence of ctDNA at the post-NAC/preoperative time point than by changes in ctDNA occurring during the surgical period and may have been affected by surgery.
When we looked at the difference in average cFAs between T1 and T2 for these 39 patients, we found a dramatic decrease after surgery by 91% (P = 0.009) for all variants and by 96% for the variants detectable at T1 (P = 0.006; Supplementary Fig. S7C). When we analyzed the different RCB groups, we found drastic decreases in cFA before and after surgery for RCB 2 (96%, P = 0.02) and RCB 3 patients (91%, P = 0.03), coinciding with the greater tumor burden before surgery in these patients, compared with RCB 1 patients (69%, P = 0.234; Supplementary Fig. S7D). Therefore, these results suggest that the extent of ctDNA decrease after surgery is proportional to the amount of tumor removed at surgery, further supporting ctDNA as a marker of tumor burden.

ctDNA status during postoperative time points predicts relapse and is prognostic
To evaluate MRD detection, we analyzed ctDNA detectability in postoperative ctDNA samples for 72 patients at T2, 55 patients at T3, and 66 patients at T4. ctDNA detection at any postoperative time point identified 24 of 26 relapsed patients (sensitivity = 92% and specificity = 21%; Supplementary Fig. S8A). The PPV for relapse detection was 36%, with 24 of 66 ctDNA+ patients experiencing relapse. Among 13 ctDNA− patients, 11 remained relapse free, resulting in an NPV of 85%. Of the two patients who relapsed, ctDNA had been detectable at the T1 time point. For one of these patients, additional samples collected after T4 revealed ctDNA detection 10 months after T4 and 10 months before clinical relapse (Supplementary Fig. S9A). In fact, the sensitivity of the postoperative test allowed for relapse detection with a median lead time before clinical relapse of 12 months (range, 1.57–38.07). The median lead time for local relapses was 15 months (n = 5; range, 6.9–32.3) versus 12 months (n = 19; range, 1.57–38.07) for distant relapses, with local relapses having somewhat lower cFA of detectable variants at the time of relapse detection (Supplementary Fig. S9B).
Patients with no detectable ctDNA at any time point postoperatively showed a trend toward improved RFS (RFS = 85%) compared with those with detectable ctDNA (58%), with an HR of 0.34 (95% CI, 0.13–0.92; P = 0.130). Similarly, there was a trend toward improved OS in ctDNA– patients (RFS 92%) compared with those with detectable ctDNA (51%), with an HR of 0.202 (95% CI, 0.066–0.618; P = 0.079), although these differences were not statistically significant (Supplementary Fig. S8B).
When analyzing each postoperative time point individually (Supplementary Fig. S8C–S8F; Fig. 4E and F), ctDNA was detectable in 71% (51 of 72) of patients at T2, 65% (36 of 55) at T3, and 59% (39 of 66) at T4. The rate of relapses tended to be higher in ctDNA+ patients across all time points but was only significant at T4, with 36% of ctDNA+ patients relapsing compared with 11% of ctDNA− patients (P = 0.0431). The PPV for relapse detection at the late postoperative T4 time point, was 36% and NPV was 89%, whereas sensitivity and specificity were 82% and 49%, respectively (Supplementary Fig. S8E).
The prognosis for patients with undetectable ctDNA (ctDNA−) was generally better than for those with detectable ctDNA (ctDNA+), across all time points, although this difference reached statistical significance only at T4. Specifically, patients with undetectable ctDNA at T4 exhibited significantly longer RFS (P = 0.015; HR, 0.24; 95% CI, 0.093–0.62) and OS (P = 0.036; HR, 0.147; 95% CI, 0.037–0.59; Supplementary Fig. S8E and S8F). Notably, all 27 patients with negative ctDNA at T4 remained alive at the time of our median follow-up of 38 months, and the 5-year OS for ctDNA− patients at T4 was 100%.
We observed that the average concentration of detectable ctDNA (cFA) was significantly higher (> fourfold) in patients who relapsed compared with those who did not relapse across all postoperative time points (P < 0.05), except at T2 (P = 0.2; Supplementary Fig. S9C). The cFA levels of the patients who relapsed at this first postoperative time point were much lower than those at T3 and T4, suggesting that ctDNA is being diluted by larger amounts of cfDNA being released as the result of trauma and healing following surgical intervention.
We analyzed separately those patients who received adjuvant capecitabine (Fig. 4A–H). A total of 67% (39 of 58) had detectable ctDNA before the initiation of treatment at T2, 65% during treatment at T3 (36 of 55), and 55% (28 of 51) after the completion of treatment at T4. Sixty-one patients had at least one postsurgery (T2, T3, and/or T4) collection available for MRD analysis; 52 of these patients were ctDNA+, whereas nine remained ctDNA– throughout the adjuvant period (Fig. 4A). When considering all postoperative time points together in patients who received capecitabine, 95% of relapsed patients were ctDNA+ (18 of 19) and the median lead time to the detection of relapse was 9.3 months (range, 1.6–32.3). Patients who received capecitabine and were ctDNA– showed a trend toward improved RFS (P = 0.15) and had significantly higher OS than ctDNA+ patients (P = 0.042; Fig. 4A and B). Survival was 100% at 5 years for ctDNA– patients compared with 54% for ctDNA+ patients. No significant association with RFS or OS was found when we analyzed each time point separately (Fig. 4C–H).

CtDNA detection is an independent prognostic biomarker
To show that ctDNA status is an independent prognostic factor, we performed a multivariate analysis on our data. We found that ctDNA detectability is an independent prognostic factor [P = 0.002; HR, 24.7 (95% CI, 3.20–191.34)] in multivariate analysis (Supplementary Fig. S10; Supplementary Table S10). Both RCB status [RCB 3: P = 0.027 (HR, 11.2; 95% CI, 1.32–95.64)] and adjuvant capecitabine [P = 0.019; HR, 0.4 (95% CI, 0.18–0.86)] are also independent risk factors in multivariate analysis in our cohort. Patients who were ctDNA+ or had higher RCB scores were more likely to relapse. Patients who received adjuvant capecitabine were less likely to relapse.

ctDNA dynamics and clearance during the adjuvant period
There were 59 patients with plasma collected at both T2 and T4 time points. Forty-two patients were ctDNA+ at T2, and of these, 14 patients (33%) became ctDNA– at the T4 collection. On the other hand, there were nine patients who were ctDNA− at T2 who became ctDNA+ at T4 and only one of these patients relapsed (Supplementary Fig. S11A). In fact, these patients had similar RFS as patients who cleared (+/−) their ctDNA (Supplementary Fig. S11B). Patients with persistent ctDNA (+/+) had significantly worse 5-year RFS (51%) than patients with consistently undetectable ctDNA (−/−; P = 0.036; Supplementary Fig. S11B). The T4 cFA for detectable variants in the −/+ patients was markedly lower (almost fivefold lower; P = 0.016) than the cFA of detectable variants in the 28 patients with persistent ctDNA (+/+; Supplementary Fig. S11C). We therefore looked at whether cFA changes between T2 and T4 were different in relapsed versus nonrelapsed patients. We found that patients who relapsed had significantly increased cFA at T4 for all variants tested (average = 0.72) compared with T2 (average = 0.10; P = 0.002) whereas nonrelapsed patients had similar cFA at both time points (0.05 at T2 and 0.06 at T4; P = 0.67; Supplementary Fig. S11D). The increasing ctDNA signal observed in patients who relapsed indicates an increasing MRD burden and highlights the potential benefit of early therapeutic intervention in these patients, at the end of adjuvant therapy.
We analyzed ctDNA dynamics in the 42 ctDNA+ patients at T2 who also had blood collected at T4; 32 of these patients received capecitabine, whereas 10 did not. In the 10 noncapecitabine patients, only one (9%) cleared and that patient did not show relapse. Of 32 patients receiving adjuvant capecitabine, 13 cleared their ctDNA by the T4 time point (41%), significantly more than in the cohort without capecitabine (P < 0.0001; Fig. 5A, E and F). Three of these 13 patients with ctDNA clearance had become negative by the T3 time point, whereas the remainder still had detectable ctDNA at T3, suggesting a continued effect of capecitabine over the 6-month treatment period in the majority of these patients. Two of the 13 patients with clearance showed disease relapse (15%) whereas the rate of relapse for patients without clearance was 37% (P = 0.10; Fig. 5B). The RFS for patients on capecitabine who cleared their ctDNA at T4 tended to be better (83%) than that for those who did not clear (56%), with HR of 0.29 (95% CI, 0.079–1.08; P = 0.10; Fig. 5B). The OS at 5 years was 100% and 52% in patients who cleared and did not clear, respectively, but it did not reach statistical significance (P = 0.20; HR, 0.26; 95% CI, 0.046–1.53; Fig. 5B).
Among the 19 patients treated with adjuvant capecitabine who did not clear their ctDNA, those who remained relapse free (n = 12) exhibited a trend toward decreased cFA from T2 to T4 (P = 0.13), whereas those who experienced relapse (n = 7) showed a significant increase in cFA (P = 0.019; Fig. 5C). In these seven relapsed patients, the cFA continuously increased by 93% from T2 to T3 and by 73% from T3 to T4 (P = 0.0007), unlike in patients who did not relapse (Fig. 5D). Therefore, our results demonstrate that there is increased ctDNA clearance in patients treated with capecitabine and that MRD tumor burden tends to decline, even in the absence of complete ctDNA clearance, in patients who remain relapse free.

Discussion
CtDNA has been used as a plasma-based biomarker to monitor treatment and tumor progression in many cancers, including breast cancer (15–19). The detection of ctDNA using tumor-informed assays has been shown to be highly sensitive for the detection of MRD in breast cancer (18, 20–22). The present study is the largest reported series of patients with TNBC with residual disease at surgery with serial ctDNA measurements before surgery (T1), early after surgery (T2), during (T3), and immediately after capecitabine treatment (T4). Using a unique approach to tumor-informed ctDNA detection in 92 patients with non-pCR TNBC, we found that our ctDNA test allows the stratification of prognostic groups in the neoadjuvant and adjuvant settings for the management of patients with non-pCR TNBC. We were able to predict 97% of relapses at a median of 12 months before clinical relapse in this poor-risk group of patients with TNBC. The very high NPV and sensitivity of our test allowed us to distinguish patients with excellent prognosis from those with poor prognosis. We observed that our test behaves with different sensitivity and specificity at each time point. Of the four time points tested, ctDNA detection at the post-NAC/preoperative time point (T1) clearly had the strongest prognostic value for both RFS and OS, reaching an RFS of 95% and an OS of 95%. In our cohort, only one ctDNA– patient (3%) at this time point showed distant disease relapse, suggesting that the lack of ctDNA detectability at T1 has sufficient prognostic power to consider proposing trials of adjuvant therapy deescalation. Our reported sensitivity for ctDNA detection at T1 is higher than that reported in several recent studies (23–25). There was a key difference in patient selection: our analysis focused only on patients with non-pCR TNBC, whereas other studies included both pCR and non-pCR cases or did not stratify by response status. The I-SPY study (25) reported ∼60% sensitivity in patients with non-pCR TNBC, lower than our results, despite using a 16-variant Natera assay. We attribute the higher sensitivity in our study to our optimized cfDNA extraction and primer-based preamplification step, which allowed reliable detection using only five patient-specific variants. We have previously shown (11) that this approach can accurately detect ctDNA at the post-NAC/preoperative time point at a similar sensitivity to the current study. The present report is a validation of that previous study in a separate cohort, which included patients with both pCR and non-pCR, strengthening our confidence in the robustness of our approach. In our previous study (11), we reported that even among seven pCR patients, none of the five ctDNA– patients relapsed, whereas the one patient who relapsed was ctDNA+. These results are among the best reported ctDNA MRD results using a single landmark time point (T1) and highlight the validity of our tumor-informed five-variant preamplification ddPCR approach, which can be performed in hospital laboratories. Interestingly, all three of the local relapses were also ctDNA+ at T1, suggesting that this test at T1 can predict local relapses and distant relapses.
In the MRD setting (once the tumor has been removed), sensitivity and specificity were less than those at the T1 time point, and the best of these postoperative time points is the T4 late postoperative/after capecitabine time point. The weaker prognostic effect of the first postoperative time point (T2) may be due to post-surgical dilution of ctDNA by cfDNA originating from traumatized and repairing tissues as described previously (26). In fact, it has been shown that there is an influx of trauma-induced cfDNA release following surgery that can persist up to 4 weeks (26). These studies suggest that a secondary plasma sample be taken after 4 weeks for patients who are initially postoperatively ctDNA–. Given the preamplification procedure in our protocol, we could not measure the absolute quantity of cfDNA in our samples. However, the amount of ctDNA (cFA) detected was several-fold lower at T2 compared with the other time points in patients who relapsed. In fact, three of the four patients who were ctDNA− at T2 and relapsed were later ctDNA+ before relapse, and all these patients had their plasma sample taken within that 4-week period after surgery.
The extent of residual disease is an important prognostic factor in patients with TNBC receiving neoadjuvant therapy. Previous studies have shown an association between ctDNA detection and RCB score (27, 28). We found that ctDNA can distinguish good prognosis from poor prognosis very well, especially in patients with more extensive residual disease, RCB 2 and RCB 3 patients. Although RCB 1 patients had the best outcomes regardless of ctDNA status at T1, RCB 2 and RCB 3 patients showed a significant association between ctDNA positivity at T1 and worse RFS and OS. Notably, all RCB 3 patients who were ctDNA− at T1 remained relapse free, whereas those who were ctDNA+, all experienced relapse, highlighting the potential of our ctDNA test to further stratify relapse risk within the RCB 3 subgroup with 100% sensitivity and 100% specificity. In fact, the significant prevalence of negative ctDNA in RCB2 (nine of 30) and RCB3 (four of 20) patients suggests that the detection of ctDNA reflects a deeper biology than merely the presence of sufficient tumor cells in the breast. The nonrelapsed RCB2 and RCB3 tumors may represent biologically indolent tumors or tumors that are not disseminating, explaining their better outcomes. In fact, in nine of these tumors for which we have TNBC type data (29), five are luminal/androgen receptor (LAR) subtype. LAR subtypes are often associated with less aggressive behavior compared with classical basal-like TNBC and are less responsive to NAC (30).
Although the RCB score only reflects response to NAC, ctDNA reflects both the local tumor burden and the presence of micrometastatic disease. These findings suggest that ctDNA detection may be superior or complementary to RCB scoring in assessing prognosis in patients with non-pCR TNBC, with ctDNA being especially clinically useful in the RCB 2 and RCB 3 groups of patients. To maximize the utility of ctDNA detection, we plan to explore composite models that combine ctDNA with RCB score and also imaging response to neoadjuvant therapy to better capture the biological heterogeneity of residual disease and more accurately identify patients who are truly at higher or lower risk.
Given the collection of plasma at four time points, we were also able to study ctDNA dynamics both before and after surgery and before and after capecitabine therapy. The impact of surgery on ctDNA levels was significant, particularly in patients with the most residual disease (RCB 3), who showed the greatest decrease in ctDNA proportions from T1 to T2. This suggests that the magnitude of ctDNA reduction after surgery reflects the volume of tumor resected, further supporting ctDNA as a dynamic marker of tumor burden. In addition, capecitabine treatment was associated with a significantly higher rate of ctDNA clearance (41% vs. 10%), with patients having cleared their ctDNA during the adjuvant period (from T2 to T4) showing excellent prognosis. In contrast, patients who relapsed showed increases in ctDNA concentration during the adjuvant period. Interestingly, the mid-adjuvant time point was not adequate to identify all cleared patients, suggesting that 3 months of capecitabine may not be sufficient to prevent relapse. These findings suggest that monitoring ctDNA dynamics during adjuvant therapy may help identify patients who could be candidates for clinical trials exploring therapeutic escalation beyond capecitabine. These findings support a prospective trial evaluating ctDNA-guided de-escalation of adjuvant capecitabine in patients with non-pCR TNBC who are ctDNA-negative at the T1 time point, with RFS as the primary endpoint. The most robust approach to establish clinical utility would be a biomarker strategy design in which these patients could be randomized to ctDNA-guided therapy versus standard-of-care capecitabine, or a more classic enrichment design randomizing ctDNA− patients to adjuvant capecitabine or no adjuvant capecitabine. However, given the excellent prognosis of ctDNA− patients and the 95% NPV for no relapse, a less costly and more feasible alternative would be a single-cohort study design, in which capecitabine is omitted in ctDNA− patients. Once clinical utility is established, this assay would be ready for implementation in real-world clinical practice. Barriers to implementation of these in-house assays include cost, turnaround time, and standardization. Our validated results with five variants show that our approach provides very high sensitivity with the added benefit of reducing cost (the price per patient for the analysis of four time points is approximately USD $4,500), thus striking a good balance between sensitivity and cost-effectiveness. As tumor sequencing costs continue to decrease, the cost of developing the assay will become more affordable; right now, its cost is comparable if not less than most commercially available ctDNA assays, especially if serial testing is proposed. Regarding turnaround time, we have now streamlined the entire process such that the assay’s turnaround time from tumor sequencing to ctDNA analysis is about 6 to 7 weeks. For transfer to other laboratories, interlaboratory analytic validation would have to be performed. However, the test is relatively simple technically, and its most challenging aspects involve DNA extraction from tumor biopsies and the performance of digital PCR, a technique presently used in many clinical laboratories.
Our approach of tumor-informed mutant variant detection has several limitations. First, all tumor-informed assays for each patient with TNBC are unique and rarely overlap with others, making the generation of these assays labor-intensive and time-consuming. Second, tumor-informed ddPCR assays cannot detect newly emerging mutations during disease progression although emergent clones are not very pertinent to this study as we are looking at prognosis at early time points. Third, there is a lack of external validation. Future external cohort data validation will require establishing multicenter collaborations or building a consortium across institutions to enable patient recruitment, prospective harmonization of sample collection, cfDNA extraction, and clinical annotation, thereby enabling access to larger and more diverse sample cohorts. In addition, data-sharing initiatives, such as shared somatic variant repositories within a federated data-sharing framework, would facilitate external validation of personalized ctDNA assays without transferring raw sequencing data and preserving patient privacy. Fourth, we do not rule out the possibility that false-positives may be introduced because of polymerase errors, as our workflow introduces a preamplification step enabling the measurement of all five variants to be performed with only 2 mL of plasma, which is relatively small quantity for routine blood collections. Our findings in the T1/T2 −/+ and T2/T4 −/+ subgroups diverged from initial expectations, raising some considerations about both the biological interpretation and technical limitations of our assay. Closer observation revealed that many of these false-positive results were due to cFA levels that were very close to the threshold. These low levels of ctDNA may also reflect residual micrometastatic disease in a dormant state, ultimately eliminated or controlled by the host immune system without progressing to overt relapse. In fact, we found that eight of these 16 patients had cleared their ctDNA at later follow-up time points. To minimize false-positives, one can increase the number of normal plasma replicates, adjust the threshold, and/or integrate with clinical and radiologic features. Adjusting the threshold of each assay is feasible, but we must consider the clinical use of the test. If the test is to be used for therapy de-escalation, we would prioritize fewer false-negatives (i.e., maximize sensitivity). If the test is to be used to escalate therapy, then we would prioritize fewer false-positives (i.e., maximize specificity). An example of adjusting the threshold to increase specificity is shown in Supplementary Fig. S12. This highlights one of the benefits of an academic-based test: we can adjust the performance of the test according to the desired indication. For the present study, the use of the proposed threshold resulted in the maximum sensitivity with the least loss of specificity.
Despite these limitations, the performance of the post-NAC/preoperative T1 time point was outstanding in predicting the absence of relapse with specificity >95%. Perhaps because of logistic difficulties in obtaining plasma samples at this time point, the post-NAC preoperative time point has only been studied in a few previous reports. One study sequenced ctDNA using a fixed panel of 275 genes and found that end-of-treatment ctDNA status is prognostic and associates with RCB score, but the performance of their test was not optimal to guide adjuvant therapy (27). More recently, a tumor-informed method using a 1,021-gene panel to guide cfDNA next-generation sequencing analysis found an association with poor outcomes in post-NAC samples but with suboptimal NPV (<90%; ref. 23). The latest-generation “ultra-sensitive” assays in which >1,000 genes are probed simultaneously have the potential to significantly improve these results (31). A fixed panel would not perform well for TNBC (as mentioned above) as there are very few recurrent mutations for this subtype. In our cohort, none of the 467 selected variants were derived from recurrently mutated genes, with the exception of TP53; however, the specific single-nucleotide variant sites differed across cases. Therefore, clinically used panels of common variants in recurrently mutated genes would not be adequate to capture the unique genomic landscape of each TNBC. The robustness of our test in the post-NAC T1 time point, as observed now in two separate TNBC studies, suggests that this personalized approach to treatment management is ready for prospective testing of its clinical utility in patients who have undergone NAC and require additional adjuvant therapy.

Supplementary Material
Figure S1Supplementary Figure S1. Patient consort diagram.

Figure S2Supplementary Figure S2. Analytical validation of the mutation detection protocol.

Figure S3Supplementary Figure S3. Prognostic performance of additional adjuvant capecitabine.

Figure S4Supplementary Figure S4. Pearson Correlation of VAF for each tumor variant from WES and its corresponding validated ddPCR assay.

Figure S5Supplementary Figure S5. Proportion of the number of variants detected per patient at each time point.

Figure S6Supplementary Figure S6. Prognostic performance of Residual Cancer Burden (RCB) Score and ctDNA detection at the post-neoadjuvant, pre-surgical time point (T1) stratified by RCB score.

Figure S7Supplementary Figure S7. Prognostic performance of the changes in ctDNA detection from the post-neoadjuvant, pre-surgical time point (T1) to the first post-surgical time point (T2).

Figure S8Supplementary Figure S8. Prognostic performance of ctDNA detection at post-surgical time points (T2 and T4).

Figure S9Supplementary Figure S9. Fractional abundance differences in the post-surgical time points.

Figure S10Supplementary Figure S10. CtDNA detection is an independent prognostic biomarker.

Figure S11Supplementary Figure S11. Prognostic performance of the changes in ctDNA detection from the first post-surgical time point (T2) to the last post-surgical time point (T4).

Supplementary Figure S12T1 RFS curves with adjusted threshold of 3 Standard Deviations to increase specificity.

Table S1Supplementary Table S1. Representativeness of Study Participants

Table S2Supplementary Table S2. Variant identification with specific primers and probes for ddPCR.

Table S3Supplementary Table S3. Assays for Validation.

Table S4Supplementary Table S4. Within run assay precision.

Table S5Supplementary Table S5. Between run assay precision.

Table S6Supplementary Table S6. Statistics on protocol validation.

Table S7Supplementary Table S7. CtDNA status and clinical outcomes of TRICIA patients.

Table S8Supplementary Table S8. Raw fractional abundance for each variant in patient plasma, normal plasma and tumor tissue.

Table S9Supplementary Table S9. Corrected Fractional Abundance for each variant.

Table S10Supplementary Table S10. Summary of results from multivariate analysis.