Colorectal Cancer Recurrence Prediction Using a Tissue-Free Epigenomic Minimal Residual Disease Assay

Colorectal cancer remains a significant global health burden, accounting for more than 900,000 annual cancer-related deaths worldwide (1). Despite advances in surgical techniques and adjuvant therapies, 20% to 30% of patients with stage II to III colorectal cancer treated with curative intent experience recurrence following initial treatment (2–5). More than 50% may be cured with surgery alone, but many are exposed to unnecessary toxicity from adjuvant chemotherapy due to the imprecision of diagnostic tools to distinguish those at risk of recurrence.

Currently, the decision to administer adjuvant chemotherapy is primarily based on clinicopathologic risk factors including stage, histologic grade, and lymph node involvement (6–9). These conventional approaches have limitations in accurately predicting disease recurrence, as they do not fully capture biological heterogeneity around each tumor’s metastatic potential or account for the presence of residual microscopic disease after treatment. After the completion of curative intent therapy, patients are monitored postoperatively every 3 to 6 months for up to 5 years, to identify early metastatic relapse amenable to curative treatment, with minimal improvement in outcomes from serial CT or carcinoembryonic antigen (CEA) measurement documented to date (10–15).

Recent advances in molecular diagnostics have opened new avenues for improving the management and surveillance of early-stage colorectal cancer. Among these, ctDNA analysis is a promising noninvasive technique for the detection of minimal/molecular residual disease (MRD) and monitoring for recurrence (16). The concentration of ctDNA is generally lower in early- versus late-stage disease, requiring highly sensitive detection methods (17, 18). Additionally, clonal expansion of hematopoietic stem cells carrying somatic mutations (“clonal hematopoiesis”) complicates accurate ctDNA-based MRD detection given its high prevalence in the general population (detectable in more than 50%–80% of adults) and because of overlap in mutation profiles at low blood concentrations (19–21). To overcome these biological challenges, techniques have been developed to optimize sensitive and specific detection of ctDNA for MRD detection in solid tumors, including deep sequencing targeting a single or small number of variants predefined by tissue sequencing, error correction tools, methylation interrogation, fragmentomic assessment, and multianalyte approaches (22, 23).

The majority of currently available MRD options for patients with colorectal cancer rely on tissue sequencing to predefine variants to track in plasma using customized, patient-specific assays. Up-front tissue analysis can be logistically complex, particularly when patient care has occurred at different centers, which may limit clinical utility for adjuvant decision-making due to long turnaround times. Furthermore, tissue availability is limited for some patients, such as in the neoadjuvant treatment setting. Here, we evaluate the clinical performance of a plasma-only epigenomic MRD assay for postoperative longitudinal recurrence prediction. This assay was designed to provide rapid turnaround of results in the postsurgery and surveillance settings with low logistical burden to expand patient access and ensure result delivery within clinically relevant decision-making time frames.

COSMOS-CRC-01 (UMIN000037765) is an ongoing multicenter prospective nonrandomized observational study in Japan. Key eligibility criteria included (i) ages ≥20 years at the time of informed consent, (ii) confirmed diagnosis of clinical stage 0 to III colorectal cancer (Union for International Cancer Control eighth edition), and (iii) planned surgical resection. Whole-blood samples were collected in Streck tubes preoperatively (for use in a separate early cancer detection study). Patients with stage I and higher colorectal cancer were eligible for this MRD study and had samples collected at day 28 and every 3 to 6 months postoperatively for up to 5 years or until clinical recurrence (Supplementary Fig. S1). Samples were collected at the time of recurrence when possible. CT was performed every 6 months after surgery according to the Japanese Society of Cancer of the Colon and Rectum (JSCCR) guidelines and thus occurred concurrently with samples collected at 6, 12, 18, 24, 30, 36, 42, 48, 54, and 60 months postresection (8). Plasma was isolated from whole blood and banked for retrospective testing. Results were blinded (not available to the treating clinician), and all therapeutic decisions followed the standard of care. The study was conducted in accordance with the Declaration of Helsinki and the Japanese Ethical Guidelines for Medical and Biological Research Involving Human Subjects. The study was approved by the Institutional Review Board of the National Cancer Center Japan and authorized by the head of each participating institution. All eligible patients provided written informed consent and were enrolled immediately after consent.

Enrollment of 501 patients (which included 159 patients with stage 0 colorectal cancer who were ineligible for this study) occurred between January 2020 and April 2021. In this interim analysis (data locked March 31, 2023), we evaluated all available postoperative samples collected as of September 30, 2022, from all enrolled patients with clinical stage I to III disease (n = 342 patients; median age, 70 years; 44% female). Two patients with R2 resection were excluded from all analysis sets given known persistent disease within the first month after surgery. Two patients with no postoperative time points were also not included. Presurgery samples eligible for colorectal cancer early-detection studies will be presented elsewhere. Clinical data, including demographics, tumor site, T and N (tumor and node) staging, clinicopathologic characteristics, timing and use of adjuvant and neoadjuvant chemotherapy, and longitudinal follow-up, were abstracted from standard-of-care electronic health records. Blood CEA levels were available for a subset of patients at the day 28 postoperative time point. Recurrence status was determined according to standard methods with diagnostic imaging or any other diagnostic procedure if imaging was not confirmative.

MRD analysis was performed using the analytically validated, commercially available tissue-agnostic Guardant Reveal assay (Guardant Health; ref. 24). Up to 30 ng of cell-free DNA extracted from 2 to 5 mL of plasma is partitioned based on its methylation state, repaired, ligated with custom adapters for molecular tracking, and PCR-amplified. The libraries are enriched for a panel covering the majority of the methylome, which includes more than 20,000 epigenomic regions that are differentially methylated in colorectal cancer and other tumor types as well as methylated control regions. The enriched libraries undergo next-generation sequencing on the Illumina NovaSeq 6000 platform. The sequencing data are analyzed using Guardant Health’s proprietary bioinformatic pipeline software for MRD detection, which is trained to detect the presence of ctDNA based on epigenomic signals. The bioinformatic algorithm for colorectal cancer evaluates ∼2,000 differentially methylated regions (DMR) most strongly associated with colorectal cancer and classifies each sample as ctDNA detected (positive) or not-detected (negative) based on a predefined statistical likelihood threshold that the patterns of methylation are tumor-derived. Tumor fraction is estimated by normalizing cancer-specific DMR with appropriately matched control regions within each sample. A methylation score is calculated for each cancer-specific DMR. The DMR with the highest scores are selected and averaged to estimate the tumor fraction.

Supplementary Figure S2 shows a flow diagram of criteria for inclusion/exclusion in the three main analyses for the study: longitudinal surveillance, 28-day postsurgery, and postchemotherapy/chemotherapy clearance. The primary endpoint of this study was ctDNA-positive rate postoperatively with longitudinal surveillance (defined as having at least two postoperative samples). The majority (98%; 334/340) of enrolled patients eligible for the MRD study were evaluable for the primary endpoint, with few excluded because of sample availability or lost to follow-up (six patients had insufficient postsurgical sampling, and all patients had at least 8.9 months of follow-up). Patients (n = 3) with early relapse (within 6 months of surgery) were included in the longitudinal analysis even if only one sample was available, as the opportunity to collect more than one longitudinal sample was limited. Three of the four patients with a single postoperative time point ineligible for the longitudinal surveillance analysis were included in the day-28 analysis. The fourth patient had only a 3-month time point (negative in a patient with stage I nonrecurrent disease).

To understand the relationship between postoperative ctDNA detection and colorectal cancer recurrence, we evaluated the recurrence-free interval (RFI) for patients with ctDNA ever detected versus never detected during postoperative surveillance and at each longitudinal time point, as well as sensitivity, specificity, and lead time from ctDNA detection to radiographic recurrence. Secondary endpoints include subgroup analyses across clinically relevant populations in the longitudinal analysis population, the association of RFI with ctDNA detection at day 28 postsurgery, and the association of RFI with adjuvant chemotherapy use. Given that only three patients with stage I disease had a recurrence or detectable ctDNA, we focused the day-28 and postchemotherapy analyses on patients with stage II and higher disease (intended-use population). For the adjuvant chemotherapy analyses, we evaluated the first available postchemotherapy sample (main analysis) as well as samples collected at the 3-month postsurgical time point for patients who received adjuvant chemotherapy (exploratory analysis, akin to an early on-treatment time point). RFI was evaluated for these single time points as well as for four subgroups (persistent negative, cleared, converted positive, and persistent positive) defined by a paired time point collected before the start of adjuvant chemotherapy (the day-28 postsurgery time point).

The Kaplan–Meier method was used to estimate RFI based on ctDNA detection status for each of the time-to-event analyses. RFI was defined as the time from surgery to clinical diagnosis of recurrence. Patients without recurrence were censored at the time of the last follow-up. For each longitudinal time point, we also evaluated RFI defined as the time from sample collection to clinical recurrence to better describe outcomes expected after each surveillance sample was collected. Differences in RFI [HR, 95% confidence interval (CI), and P value] were measured using the log-rank test. Given the potential utility for the day-28 postsurgery time point to inform adjuvant decision-making, we evaluated the correlation between RFI and ctDNA detection status using Cox proportional hazard modeling in patients with stage II or higher disease to confirm the strength of ctDNA detection status while controlling for other prognostic clinicopathologic factors. Covariates included patient age, gender, T and N stage, tumor location, use of adjuvant chemotherapy, lymphovascular and perineural invasion, and the corresponding CEA level at day 28 postsurgery. Although microsatellite instability status was known, it could not be included as a covariate because none of the 18 patients with high microsatellite instability eligible for the day-28 time point analysis (one of whom had recurrence) had ctDNA detected at the day-28 time point. Tissue genotyping for RAS and RAF mutations is not routinely performed and was not available to include as covariates at the time of this analysis. We also evaluated RFI within each pathologic stage (II or higher) for the day-28 time point.

The target recruitment of 500 patients was the maximum achievable within the available funding and the recruitment time frame, and thus, the sample size of this descriptive study was not predefined based on a statistical power calculation. Statistical analyses were performed using GraphPad Prism version 10.2 (GraphPad Prism; RRID: SCR_002798).

Processed data (ctDNA detected or not-detected results) and patient-level clinical data generated in this study are available within the article Supplementary Materials. Raw sequencing data for this study were generated at Guardant Health and are not publicly available to ensure compliance with patient consent forms, but further details can be made available through the corresponding author upon reasonable request.

A total of 342 patients with clinical stage I to III colorectal cancer were enrolled in the multicenter observational study COSMOS-CRC-01. Clinicopathologic characteristics are summarized in Supplementary Table S1. The population included a representative breakdown of patients with both colon and rectal cancer (65% and 35%, respectively). The majority (62%) had pathologic stage II or III disease. Six patients with synchronous primary colorectal cancer were categorized according to the stage of their more advanced tumor. We included 13 patients with clinical stage III disease that had pathologic stage IV disease following surgical resection, eight of whom were classified as stage IV due to nodal involvement classified as N3 by Japanese guidelines but not considered to be regional by Union for International Cancer Control guidelines. Nine patients who underwent endoscopic resection of a stage I tumor, for which the resection status could not be confirmed, were included.

At data cutoff, 89% of subjects had at least 24 months of follow-up from surgery (median, 28.4 months; range, 8.9–38 months). Forty-seven patients (14%) experienced recurrence, including both patients with R2 resection (both were ctDNA-positive after surgery and excluded from all analyses). Recurrences consisted of a single metastasis to the lung (n = 17, 36%), liver (n = 12, 25%), peritoneum (n = 5, 11%), locoregional sites (n = 4, 8%), central nervous system (n = 1, 2%), or multiple sites of metastasis (n = 8, 17%). Ten patients (3%) died during the follow-up period, and 11 (3%) developed a second primary cancer.

A total of 1,977 eligible postoperative longitudinal samples (collected at least 21 days after definitive resection) were available from 338 patients with cancer who had undergone R0/R1 or endoscopic resection (median, 6 per patient; range, 1–8), of which 1,906 were successfully analyzed [97% sample quality control (QC) pass rate]. The median cell-free DNA input of samples passing QC was 18.8 ng (IQR, 13.1–27.2 ng). All clinical and testing data can be found in Supplementary Table S2.

The longitudinal surveillance analysis was conducted on 1,902 postoperative samples passing QC from 334 patients (Supplementary Fig. S3A–S3C). ctDNA was detected in at least one sample in 51 patients (15%), including 3/115 (3%) patients with stage I, 13/99 (13%) stage II, 28/108 (26%) stage III, and 7/12 (58%) patients with stage IV disease. A total of 42 patients remained positive after completion of all treatments. Seven patients cleared with chemotherapy without repositivization, one was positive only during adjuvant chemotherapy, and one patient with recurrence had no additional posttreatment samples. The 24-month RFI for the 292 patients who remained ctDNA-negative posttreatment was significantly higher than that in the 42 patients who had a posttreatment ctDNA-positive test result (94.7% vs. 40.7%; HR, 16.70, 95% CI, 5.68–49.09; P < 0.0001; Fig. 1A). Among patients with ctDNA detected before clinical detection of recurrence, the median lead time was 5.3 months (IQR, 3.0–16.4 months), with a maximum lead time of 28.7 months (Fig. 1B; Supplementary Fig. S3A). Clinical features associated with ctDNA detection included higher T and N stage and the presence of lymphovascular invasion (Fig. 1C). The estimated tumor fraction generally increased with sampling closer to the time of clinical detection of recurrence (Supplementary Fig. S4).

The median number of DMR detected per positive sample in this study was 994 (IQR, 792–1,203). Figure 2A and B illustrate methylation patterns observed across ∼1,700 common DMR in patients without and with recurrence, respectively.

Specificity, calculated from 1,461 posttreatment samples from 290 patients without recurrence, was 98.2% (95% CI, 97.3%–98.9%), and it was 99.1% (855/863; 95% CI, 98.2%–99.6%) for the subset of samples with at least 16.4 months of follow-up (the upper IQR of the lead time observed in this study, Supplementary Fig. S3B and S3C). Twenty-seven posttreatment samples from 16 patients who had not developed recurrence during the follow-up period had a positive result. Seven of these 16 patients (44%) had multiple positive samples or a single positive sample at their last visit and may have insufficient follow-up (2.7–15 months) to confidently rule out recurrence given the long lead times observed in this study. In one patient with stage IIA disease, ctDNA was detected in three posttreatment samples, leading to a diagnosis of a second primary lung cancer, which cleared after resection of the lung tumor, and likely represents true-positive ctDNA detection of malignancy (Supplementary Fig. S3C). Two patients (one stage I and one stage IIA) had ctDNA detectable at day 28 postresection but undetectable in all subsequent samples, which may represent false-positive results; however, we cannot rule out the possibility of “self-clearance” of postsurgical MRD through immune surveillance in the postoperative period. The remaining six patients had transient positive results in one or two samples. These may represent false-positive results, although we cannot rule out the possibility that these patients have low-volume MRD near the assay limit of detection with stochastic sampling contributing to subsequent negative results. We observed a seventh patient convert from positive to negative to positive without intervening treatment before a lung relapse in a low tumor fraction setting (COSMOS0436; Supplementary Fig. S4). Details on sensitivity and specificity by stage are provided in Supplementary Table S3.

In the stage II and higher intended-use population, we observed a statistically significant difference in RFI based on ctDNA detection at each postoperative longitudinal time point evaluated in this interim analysis and a high negative predictive value (NPV), >91%, for recurrence within 1 year of each sample collection (Supplementary Fig. S5).

Sensitivity for recurrence detection increased with longitudinal sampling and was 81% (17/21, 95% CI, 58.1%–94.6%) for colon cancer and 60% (12/20, 95% CI, 36.1%–80.9%) for rectal cancer (Fig. 3A). Sensitivity was 100% among patients with liver metastases, whereas patients with recurrence limited to the lung or peritoneum had a lower overall sensitivity of 53% and 40%, respectively (Fig. 3B). The sensitivities observed are similar to three published studies of tissue-informed ctDNA assays in resected colorectal cancer that report longitudinal ctDNA detection by site of metastasis (Fig. 3C; refs. 25–27).

The proportion of patients with recurrence limited to the lungs was significantly higher for those with rectal cancer than with colon cancer (55% vs. 19%; P = 0.02; Fig. 3D). Given that the assay’s sensitivity for metastasis detection in any specific organ was similar regardless of primary tumor site (Fig. 3E), the lower longitudinal sensitivity for rectal cancer is likely due to the distribution of metastatic sites rather than differences in the ability of the assay to detect rectal versus colon cancer.

In most patients (83%) with metastases limited to the liver, ctDNA was detected at either the day-28 postsurgery time point or before clinical recurrence. In contrast, patients with recurrence limited to the peritoneum or lung had a higher proportion of cases in which ctDNA was detectable only at the time of clinical recurrence. The median time from surgery to the detection of lung metastasis was significantly longer than that for liver metastasis (17.9 vs. 6.3 months; P = 0.03), which might explain the lower proportion of patients with lung metastases with detectable ctDNA at the 28-day postsurgery time point (Fig. 3F). The likelihood of recurrence by site of metastasis and ctDNA detection status during longitudinal surveillance is described in Supplementary Fig. S6A and S6B. Among patients with recurrence, the most common sites in those with ctDNA-positive test results included the liver followed by the lung, whereas the most common sites in patients with ctDNA-negative test results included the lung and peritoneum. Despite the lower sensitivity for lung and peritoneal metastases, patients whose test results remained ctDNA-negative posttreatment had low rates of recurrence 24 months after surgery: 4.6% for lung and 7.4% for any metastasis.

The RFI analysis based on ctDNA detection at 28 days postsurgery included 216 patients (141 patients with colon cancer and 75 patients with rectal cancer) with stage II or higher disease and a sample collected 21 to 60 days postsurgery and before starting adjuvant chemotherapy, if given. The median time from surgery to sample collection was 27 days (IQR, 24–32 days), and 116 patients (53.7%) received subsequent adjuvant chemotherapy. Postsurgery blood CEA levels were available for 176 patients included in the day-28 postsurgery analysis (median, 1.8 ng/mL; IQR, 1.2–2.7).

Detection of ctDNA was the strongest predictive factor for RFI on multivariable analysis (Fig. 4A). The median RFI for the 25 patients with detectable ctDNA was significantly shorter than that for patients with undetectable ctDNA (21.6 months vs. not reached; Fig. 4B; Supplementary Table S4A). The estimated 24-month RFI rates were 46.9% for patients with detectable ctDNA compared with 88.3% for patients with undetectable ctDNA (HR, 6.48; 95% CI, 2.12–19.7). Similar trends were seen for patients with both colon and rectal cancer (Fig. 4C and D; Supplementary Table S4B and S4C). We also confirmed statistically significant differences in RFI based on day-28 ctDNA detection within stage II and III disease, which provides further refinement of risk compared with standard pathologic staging alone (Supplementary Fig. S7A–S7D; Supplementary Table S5A–S5D). There were too few patients with stage IV disease to power a conclusive analysis.

The day-28 postsurgery sensitivity was higher for patients with colon cancer than that for patients with rectal cancer [45% (10/22) vs. 26% (5/19)], which is likely related to differences in metastatic site distribution and longer time between surgery and recurrence in patients with rectal cancer (median, 18.7 vs. 11.6 months). When time-bounding for recurrences within 1 year of the day-28 sample, sensitivity was 43% in both colon and rectal cancers (6/14 and 3/7, respectively). The RFI for patients with ctDNA-negative test results was nearly identical for those with colon and rectal cancers at 12 months postsurgery (∼94%); however, there was a difference by 24 months of follow-up (90.2% vs. 81.4%; HR, 2.29; P = 0.03), further illustrating the challenge of detecting late relapses with a single time point collected after surgery. The likelihood of recurrence by the site of metastasis and ctDNA detection status at the day-28 time point is outlined in Supplementary Fig. S8A and S8B. Even with relatively modest sensitivity at day 28 postsurgery, the estimated risk of recurrence 24 months postsurgery in patients with ctDNA-negative test results was approximately 13%.

A total of 112 patients with stage II disease or higher (58% patients with colon cancer and 42% patients with rectal cancer) had paired samples available at the day-28 postoperative time point (prechemotherapy) and following adjuvant chemotherapy but before recurrence. The median duration of chemotherapy was 3.7 months (range, 2–9.6 months), and the median time of sample collection after completing chemotherapy was 53 days (range, 0–154 days). ctDNA was detected in 7.1% (8/112) of patients following chemotherapy and was associated with shorter RFI (median, 14.9 months vs. not reached; HR, 11.58; 95% CI, 1.33–101.0; P < 0.0001; Fig. 5A; Supplementary Table S6A). The rate of ctDNA positivity was similar in patients with colon and rectal cancers [7.7% (5/65) and 6.4% (3/47), respectively] as was the 24-month RFI for patients with ctDNA-negative test results (colon: 90%, 95% CI, 79.1%–95.4%; rectal: 83.5%, 95% CI, 68.5%–91.8%; P = 0.14 by the log-rank test).

The pattern of ctDNA detection before and after chemotherapy may provide additional prognostic value compared with information from a single postchemotherapy time point (Fig. 5B; Supplementary Table S6B). Although ctDNA positivity after chemotherapy was associated with poor RFI regardless of prechemotherapy ctDNA status, patients who were persistently negative at both time points had significantly improved RFI compared with patients who cleared ctDNA (HR, 0.22; 95% CI, 0.03–1.52; P = 0.0041). Forty percent of patients with initial ctDNA clearance had ctDNA detection at a later time point compared with only 5% of patients who consistently tested negative for ctDNA. RFI at 24 months postsurgery was numerically greater in patients who cleared ctDNA versus those who did not; however, few patients were available to power a statistical analysis. We did not observe a relationship between the duration of adjuvant chemotherapy and the likelihood of ctDNA clearance, noting we have few patients to power a robust analysis.

We also explored ctDNA dynamics shortly after starting adjuvant chemotherapy in the subset of patients who had received adjuvant treatment with a 3-month postsurgery time point available (n = 105). The median time since chemotherapy started at this time point was 42 days. We observed similar significant differences in RFI based on whether ctDNA was detected and that patients with persistent negative ctDNA had >87% RFI at 24 months postsurgery (Supplementary Fig. S9A and S9B; Supplementary Tables S7A and S7B). Patients who cleared at this time point had improved RFI compared with those who did not clear with a 24-month RFI of 60% (95% CI, 25.3%–82.7%) versus 0% (HR, 0.30; 95% CI, 0.07–1.32; P = 0.04). One patient (0312) had transient detection of ctDNA at month 3 that subsequently cleared, and this patient did not recur (Supplementary Fig. S3C).

This is the first study validating the clinical performance of the tissue-free, epigenomic-based Guardant Reveal MRD assay in early-stage colorectal cancer. The assay demonstrated statistically significant and clinically relevant recurrence prediction at each postsurgical time point evaluated for patients with stage II or higher colorectal cancer with a high sample pass rate (97%) and 100% patient evaluable rate. The assay showed promise for use in a small number of patients with stage IV resectable cancer, although larger studies are needed.

Our data show that ctDNA positivity at a single time point 28 days postsurgery or after adjuvant chemotherapy has a significant prognostic value that may support adjuvant and postadjuvant therapy decision-making and that longitudinal testing increases sensitivity for detection of recurrence, with overall performance similar to that of tissue-informed MRD assays. Similar to Reinert and colleagues (26) and Henriksen and colleagues (27), we highlight that ctDNA sensitivity depends on the site of recurrence, an important consideration for interpreting ctDNA results for patient care and for comparing clinical performance of ctDNA-based MRD assays across different studies.

Our group previously described lower ctDNA shedding from lung and peritoneal metastases in advanced colorectal cancer; this biological factor contributes to reduced ctDNA detection at these sites and poses a challenge for all ctDNA-based MRD assays regardless of methodology (28). In peritoneal disease, direct seeding into the peritoneal cavity, bypassing the circulatory system, may explain the lower rate of ctDNA detection in plasma (29, 30). Less is understood about ctDNA shedding from lung metastases. In primary non–small cell lung carcinoma, variable rates of ctDNA shedding have been attributed to tumor size, histology, anatomic location, metabolic activity, and underlying genotype; and undetectable ctDNA has been associated with a more favorable prognosis (31, 32). Patients with colorectal cancer with isolated lung metastases may have a better prognosis than those with other sites of metastases (33–35). Smaller, more indolent, and/or favorable prognostic tumors may have lower ctDNA shedding into the blood; however, further follow-up is needed to confirm if ctDNA detection status at the time of an isolated lung metastasis detection could have prognostic significance.

The sensitivity at the day-28 time point is modest given the challenge in detecting later recurrences, particularly after 12 months but should be interpreted in the context of the NPV, as many patients with colorectal cancer are already cured with surgery alone. We found that ctDNA negativity at day 28 translates to low 24-month postsurgery rates of recurrence, and these rates are even lower if ctDNA remains negative with longitudinal sampling. We demonstrated high NPV at each longitudinal time point; >91% of patients with ctDNA-negative results were recurrence-free within 1 year of sample collection, suggesting that serial negative ctDNA testing identifies a subset of patients with a low risk of recurrence and may be considered for risk-adapted imaging approaches. However, NPV may be lower among patients at greater risk of lung metastases, such as patients with rectal cancer, or peritoneal metastases, such as those with T4, node-positive, perforated, right-sided, mucinous, or signet-ring cell cancers or KRAS/BRAF-mutated tumors (29, 36). This is an important consideration for individualizing surveillance imaging strategies, adjudicating radiographic findings during routine surveillance, and making adjuvant chemotherapy decisions. Although several studies have suggested limited benefit of adjuvant chemotherapy among patients with postoperative ctDNA-negative results, only one included a substantial proportion of patients with rectal cancer (37–39). The relative benefit of adjuvant chemotherapy, particularly in those with high-risk features, is still being investigated in randomized controlled trials (40). Patients with rectal cancer are underrepresented, or their outcomes are not specifically reported in most studies of ctDNA MRD that focus on stage II/III resectable colorectal cancer; only one of two studies that focused on rectal cancer reported sensitivity based on the site of metastases (25–27, 37, 39, 41–47).

The variable ctDNA sensitivity by site of recurrence is also important to consider when attempting cross-study comparisons. In this study, we had a clinically representative cohort of patients with both colon and rectal cancers (65%/35%), similar to the proportion reported in the JSCCR registry of resected colorectal cancer (61%/39%; ref. 8). The proportion of metastases to the liver, lung, and peritoneum reported in the JSCCR registry was 45%/24%/16% and 30%/35%/5% for colon and rectal cancers, respectively, compared with 33%/19%/19% and 25%/55%/5% for colon and rectal cancers in this study. If the rectal cancer metastatic site distribution in this study was similar to the JSCCR registry estimated rates (i.e., 35% vs. 55% of all metastases), the longitudinal sensitivity in rectal cancer would be expected to be 72% versus the observed 60%. To further illustrate this point, if the test used in this study was applied to a theoretical cohort with a metastatic site distribution similar to that observed in patients with stage IV colon adenocarcinoma reported by surveillance, epidemiology, and end results (73% liver metastases, 21% lung metastases, and 6% multiple/other), the expected surveillance sensitivity in colon cancer would be 88% instead of the observed 81% (35). In other words, the sensitivity estimate of the same test will vary when applied to different populations with different metastatic site distributions. Furthermore, estimates of sensitivity, specificity, positive predictive value, NPV, and lead time of an assay will also depend on the cohort stage distribution, duration of follow-up, use of adjuvant chemotherapy, and frequency of blood sampling and imaging, which should also be considered if attempting cross-study comparisons.

The patients in this study received frequent radiographic imaging consistent with Japanese guidelines (similar or more frequent compared with US and European guidelines), and a median lead time of 5.3 months between ctDNA detection and clinical detection of recurrence was observed (6, 8, 48). In several cases, long lead times of 1 to 2 years were observed, which may result in an underestimation of specificity due to the need for extended follow-up. Specificity estimates were higher with longer follow-ups (99.1% vs. 98.2% when not adjusting for follow-up). Less frequent surveillance sampling in the second year of follow-up compared with the first year (every 6 months vs. every 3 months) could impact the ability to observe lead times in recurrences happening after the first year (particularly relevant for lung and peritoneal recurrences). We were not able to review previous radiographic scans to confirm if, in hindsight, there was any evidence of occult recurrence at the time of initial radiologic review. This limitation should be considered when designing MRD-directed adjuvant treatment trials, as shorter lead times have been observed in prospectively tested cohorts (49, 50).

MRD detection through ctDNA analysis is becoming a routine clinical tool that has the potential to revolutionize the management of early-stage colorectal cancer for both adjuvant therapy decision-making and recurrence monitoring. Several prospective randomized trials are ongoing to evaluate the use of ctDNA for adjuvant decision-making (40). While waiting for additional data from ongoing randomized trials to definitively establish the safety and efficacy of specific adjuvant therapy escalation and de-escalation approaches, there is still a role for clinical MRD testing in colorectal cancer to inform prognosis. ctDNA testing may help determine adjuvant treatment in patients for whom the risk and benefit of adjuvant therapy are unclear or aid in adjuvant therapy decisions in which patient preference is not aligned with standard recommendations. Furthermore, ctDNA testing may assist in postoperative surveillance to enable tailored radiographic strategies and early identification of recurrence. As a tissue-free assay, this may offer logistical advantages by reducing the burden of tissue handling on health care systems as well as improved turnaround time for the initial result during the adjuvant decision-making window while maintaining performance comparable to tissue-informed approaches. Tissue-free assays may also be the only option for patients with limited tumor specimens. These factors may be useful for both busy clinical practices and in the conduct of clinical trials to maximize the patient evaluable rate, particularly within constricted time frames for adjuvant therapy decision-making.

Y. Nakamura reports personal fees from Guardant Health Pte Ltd and Guardant Health Japan Corp, grants from Guardant Health AMEA, Inc and Guardant Health during the conduct of the study, personal fees from Natera, Inc, Roche Ltd, Premo Partners, Inc, Takeda, Exact Sciences, Gilead Sciences, MSD K.K., Eisai, Zeria Pharmaceutical, Miyarisan Pharmaceutical, Merck, Carenet, Inc, Hisamitsu Pharmaceutical, Taiho Pharmaceutical, Becton, Dickinson and Company; grants and personal fees from Seagen, Inc, Daiichi Sankyo Co Ltd, and Chugai Pharmaceutical; and grants from Genomedia, Tempus, and Roche Diagnostics K.K. outside the submitted work. N. Matsuhashi reports grants and personal fees from Abbott, Asahi Kasei Pharma, Chugai Pharmaceutical, Covidien Japan, Eli Lilly Japan, Eisai, Johnson & Johnson, Kaken Pharmaceutical, Kyowa Kirin, Terumo, and Tsumura; personal fees from AstraZeneca, Bayer Yakuhin, Bristol Myers Squibb, EA Pharma, Gunze Medical Limited, MC Medical, Merck BioPharma Japan, Miyarisan Pharmaceutical, Novartis, Olympus Marketing, Inc, Stryker, Takeda Pharmaceuticals, Viatris, and Yakult Honsha; grants, personal fees, and other support from Daiichi Sankyo; other support from EP‐CRSU, EPS Corporation, and ShiftZero K.K.; personal fees and other support from MSD and Ono Pharmaceutical; grants from Nippon Kayaku, Otsuka Pharmaceutical, and Toray Medical; and grants, personal fees, and nonfinancial support from Taiho Pharmaceutical outside the submitted work. E. Oki reports personal fees from Chugai Pharmaceutical, Bristol Meyers Squibb, Ono Pharmaceutical, Eli Lilly and Company, and Takeda Pharmaceutical and grants from Guardant Health, Inc outside the submitted work. M. Goto reports personal fees from Tsumura & Co, Daiichi Sankyo Company, Limited, Ono Pharmaceutical Co Ltd, and MSD K.K.; grants and personal fees from Taiho Pharmaceutical; and grants from Chugai Pharmaceutical and Nippon Kayaku outside the submitted work. Y. Kagawa reports personal fees from Chugai, Taiho, Ono, Merck, Takeda, and MSD outside the submitted work. T. Ohta reports personal fees from Bristol Myers Squibb Japan, Novartis AG, Daiichi Sankyo Company, Limited, EA Pharma Co, Ltd, Eli Lilly Japan K.K., Merck & Co, MSD K.K., Ono Pharmaceutical Co, Ltd, Otsuka Pharmaceutical Co, Ltd, Taiho Pharmaceutical Co, Ltd, Takeda Pharmaceutical Company Limited, AstraZeneca, and Yakult Honsha outside the submitted work. H. Bando reports personal fees from Eli Lilly Japan, Taiho Pharmaceutical, and Ono Pharmaceutical outside the submitted work. T. Yoshino reports grants and personal fees from Chugai Pharmaceutical, Takeda Pharmaceutical, Ono Pharmaceutical, and MSD K.K.; personal fees from Merck Biopharma, Bayer Yakuhin, and Sumitomo Corp; and grants from Amgen, Bristol Myers Squibb, Daiichi Sankyo, Eisai, Falco Biosystems, Genomedia, Medical & Biological Laboratories, Merus N.V., Molecular Health GmbH, Nippon Boehringer Ingelheim, Pfizer, Roche Diagnostics, Sanofi, Sysmex, and Taiho Pharmaceutical outside the submitted work. No disclosures were reported by the other authors.

Y. Nakamura: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. Y. Tsukada: Resources, investigation, writing–review and editing. N. Matsuhashi: Resources, investigation, writing–review and editing. T. Murano: Resources, investigation, writing–review and editing. M. Shiozawa: Resources, investigation, writing–review and editing. Y. Takahashi: Resources, investigation, writing–review and editing. E. Oki: Resources, investigation, writing–review and editing. M. Goto: Resources, investigation, writing–review and editing. Y. Kagawa: Resources, investigation, writing–review and editing. A. Kanazawa: Resources, investigation, writing–review and editing. T. Ohta: Resources, Investigation, writing–review and editing. A. Ouchi: Resources, investigation, writing–review and editing. H. Bando: Resources, investigation, writing–review and editing. H. Uchigata: Resources, investigation, writing–review and editing. C. Notake: Resources, investigation, writing–review and editing. H. Ikematsu: Resources, investigation, writing–review and editing. T. Yoshino: Conceptualization, resources, supervision, investigation, writing–review and editing.

The authors thank all the patients and their families who participated in this study. We are grateful to all investigators, research nurses, study coordinators, and research staff members at all the study sites who made this work possible. This work was supported by Guardant Health. We also thank the following employees of Guardant Health: Thereasa Rich and Kimberly Banks for medical writing and analytic and editorial support and Shile Zhang and Mingyang Cai for bioinformatics support. Microsatellite instability testing was supported by FALCO Biosystems.