A Tumor-Naïve ctDNA Assay Detects Minimal Residual Disease in Resected Stage II or III Colorectal Cancer and Predicts Recurrence: Subset Analysis from the GALAXY Study in CIRCULATE-Japan

Abstract

Purpose:

Analysis of ctDNA may enable early identification of patients likely to relapse, presenting an opportunity for early interventions and improved outcomes. Tumor-naïve plasma-only approaches for minimal residual disease (MRD) assessment accelerate turnaround time, enabling rapid treatment decisions and ongoing surveillance.

Experimental Design:

Plasma samples were obtained from 80 study participants with stage II or III colorectal cancer selected from CIRCULATE-Japan GALAXY. MRD status was assessed using a tumor-naïve ctDNA assay (xM) that integrates methylation and genomic variant data, delivering a binary call. MRD was assessed at 4 weeks postsurgery [landmark time point (LMT)] using methylation and genomic variant data and longitudinally (median, 22.1 months) using only methylation data.

Results:

At LMT, 69/80 study participants were evaluable (36 recurrent; 33 nonrecurrent). Of recurrent study participants, 22/36 had detectable ctDNA (MRD-positive) at LMT and 29/33 nonrecurrent study participants had undetectable ctDNA (MRD-negative), yielding a clinical sensitivity of 61.1% and specificity of 87.9%. Additionally, 74 study participants were evaluable for longitudinal performance with a clinical sensitivity of 83.3% and specificity of 89.5%. The median lead time from the first MRD-positive result to recurrence was 4.77 months overall, and 5.30 months for study participants with no adjuvant treatment. At 12 weeks postsurgery, MRD status strongly correlated with disease-free survival (adjusted HR, 9.69), outperforming carcinoembryonic antigen correlation (HR, 2.13).

Conclusions:

This tumor-naïve MRD assay demonstrated clinically meaningful performance at LMT and longitudinally, accurately predicting clinical recurrence. MRD status was a stronger prognostic biomarker for disease-free survival compared with standard-of-care carcinoembryonic antigen.

Translational Relevance.

Monitoring ctDNA is a noninvasive way of measuring minimal residual disease (MRD), identifying clinically actionable mutations at a landmark time point, and informing treatment options. Here, we present the clinical validation of xM, a liquid-based tumor-naïve MRD assay, in study participants with resected stage II and III colorectal cancer. At landmark time point, xM outputs a binary MRD-positive/negative call based on both methylation and genomic variant information, whereas longitudinal surveillance utilizes methylation data alone. xM model training on Tempus’ real-world multimodal database enables corrections for confounders such as germline variants and clonal hematopoiesis without sacrificing clinical sensitivity. Taken together, these novel aspects of the xM assay can bring forward clinically relevant biomarkers from a noninvasive blood draw.

Introduction

Colorectal cancer is the second leading cause of cancer death in the United States and the third most common in the world. Although progress has been made in the early detection and treatment of colorectal cancer, approximately 80% of patients with stage II and only 50% of patients with stage III colorectal cancer are cured by surgery alone (1, 2). Adjuvant chemotherapy (ACT) is considered standard of care for stage III and recommended for stage II if certain risk factors are present [MOSAIC trials: (3) and (4); (5–7)]. However, there is disagreement over risk factors, and studies have shown that there is great variability in outcomes even among those treated with ACT (6, 8).

Minimal residual disease (MRD) describes the small number of cancer cells left in the body after surgical resection or other treatment. Current guideline-directed MRD monitoring in colorectal cancer after curative treatment includes physical examination, colonoscopy, and serum carcinoembryonic antigen (CEA) testing to detect recurrent or residual disease. CEA is not a good marker for all types of colorectal cancer (9), and CEA levels may be elevated in patients with noncancerous conditions such as inflammatory bowel disease (10). Furthermore, levels of CEA do not directly correlate to metastasis or poor prognosis. CEA assays therefore have high specificity, but poor sensitivity, with sensitivity estimates ranging from below 20% to 100% in colorectal cancer (11, 12).

A major limitation of current guideline-directed residual disease-monitoring tests is that they cannot detect a signal from small numbers of residual cancer cells. However, ctDNA-monitoring assays are a noninvasive way to assess MRD from peripheral blood samples, even if original tumor cells are not present [reviewed (13, 14)]. Recent clinical trials of ctDNA-monitoring of colorectal cancer have suggested numerous benefits in treatment and prediction of disease recurrence. Plasma ctDNA identified MRD in patients with localized colorectal cancer and monitoring ctDNA levels postoperatively predicted recurrence in 85.5% of patients with a positive ctDNA test postchemotherapy (15). Another small-scale study found that patients with detectable ctDNA at 4 weeks postcurative surgery were seven times more likely to relapse than those with no detectable ctDNA, and those with detectable ctDNA after ACT were 17 times more likely to have a recurrence (16). Furthermore, although ctDNA tests follow a similar schedule to CEA assays, elevated ctDNA levels have been shown to occur significantly earlier than elevated CEA levels (17).

Another advantage of ctDNA-based MRD monitoring is its potential to inform treatment decisions and avoid unnecessary or ineffective therapies. Studies have also shown that the presence of ctDNA is not only prognostic of disease recurrence but can also be used to inform the administration of adjuvant therapies (18, 19). Although the presence of ctDNA can identify patients who may benefit from ACT, the absence of ctDNA postcurative resection may contraindicate further intervention (20). Another advantage is that this may also preclude unnecessary treatment costs. There are currently at least nine ongoing clinical trials, including PEGASUS, CIRCULATE-NA, BESPOKE, DYNAMIC-II, CORRECT-MRD II, ERASE-CRC, and CIRCULATE-Japan, that aim to further explore the benefit of MRD in colorectal cancer treatment (18, 21–27). A major caveat is that most MRD tests on the market are tumor-informed, meaning that they require access to tissue, DNA, or sequencing data from the original tumor sample and may be limited by DNA quality and quantity or difficult to obtain due to logistical challenges such as patients moving treatment facilities throughout their treatment.

Here, we report the clinical validation of Tempus xM (xM), a tumor-naïve MRD assay that integrates both methylation and genomic variant data to output a single binary value: MRD-positive (MRD⁺) or MRD-negative (MRD⁻). At a designated landmark time point (LMT) defined as 4 weeks after curative intent surgery, xM was able to identify MRD in a population of patients with stage II or III colorectal cancer, with a clinically meaningful correlation to disease-free survival (DFS). Furthermore, continued longitudinal testing demonstrated that ctDNA status was a stronger prognostic biomarker to DFS than standard-of-care CEA testing.

Materials and Methods

Cohort selection

For the clinical validation pilot, 80 study participants were selected from GALAXY, an observational arm of the CIRCULATE-Japan study (UMIN000039205; Table 1). The GALAXY study is a prospective large-scale registry in Japan and Taiwan designed to monitor ctDNA status for study participants with stage II to IV colorectal cancer who are candidates for complete surgical resection. The study screens study participants for ctDNA-based MRD status leading to the assignment of study participants to one of the two randomized ctDNA-guided phase III interventional trials, ALTAIR (treatment escalation; ref. 28) and VEGA (treatment de-escalation; ref. 27). This study was conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all study participants prior to participation in GALAXY. The protocol was approved by the Institutional Review Board of the National Cancer Center Japan and authorized by each participating institution (20, 29).

Table 1.

Cohort characteristics.

	Overall, n = 80	xM (methylation and variant)			P valuea
		MRD+, n = 26	MRD−, n = 43	Invalid (LMT), n = 11
Age at enrollment					0.3
Mean (SD)	67.65 (11.68)	70.85 (9.74)	68.07 (11.78)	58.45 (11.70)
Median	70	72	70	60
IQR	59.75–74.25	65.25–75.75	59.00–74.50	52.50–63.00
Min/max	37/89	44/89	41/89	37/80
Sex					0.3
Female	29 (36.3%)	6 (23.1%)	17 (39.5%)	6 (54.5%)
Male	51 (63.8%)	20 (76.9%)	26 (60.5%)	5 (45.5%)
Cancer stage					0.8
Stage 2	32 (40.0%)	10 (38.5%)	19 (44.2%)	3 (27.3%)
Stage 3	48 (60.0%)	16 (61.5%)	24 (55.8%)	8 (72.7%)
MSI-high status					0.3
Positive	8 (10.0%)	1 (3.8%)	6 (14.0%)	1 (9.1%)
Negative	72 (90.0%)	25 (96.2%)	37 (86.0%)	10 (90.9%)
RAS status					0.13
Positive	37 (46.3%)	16 (61.5%)	17 (39.5%)	4 (36.4%)
Negative	43 (53.8%)	10 (38.5%)	26 (60.5%)	7 (63.6%)
BRAF mutation status					0.08
Positive	8 (10.0%)	0 (0.0%)	7 (16.3%)	1 (9.1%)
Negative	72 (90.0%)	26 (100.0%)	36 (83.7%)	10 (90.9%)
ACT					0.4
Yes	32 (40.0%)	7 (26.9%)	17 (39.5%)	8 (72.7%)
Single (capecitabine)	1 (1.3%)	0 (0%)	0 (0%)	0 (0%)
Doublet (cape-ox)	31 (38.8%)	7 (26.9%)	7 (26.9%)	8 (72.7%)
No	48 (60.0%)	19 (73.1%)	26 (60.5%)	3 (27.3%)
Median follow-up (in months)	11.7	7.1	11.8	22.1	0.18

^a P value tests for difference between MRD⁺ and MRD⁻ groups; the invalid group is excluded from this analysis.

For initial clinical validation of the xM assay, we focused on postsurgical ctDNA detection at an LMT defined as 4 weeks after curative surgery as well as longitudinal time points in pathologic stage II or III colorectal cancer study participants. Those 80 study participants met prespecified eligibility criteria including pathologic stage II or III colorectal cancer, R0 resection, at least 1-year postsurgical resection, and a biobanked specimen available from the LMT. Of the 80 patients, two had stage IIA rectal cancer, with a primary site in the upper rectum. The only exclusion criteria were any major protocol deviations from the GALAXY study. The definition of recurrence was evidence of recurrent disease detected on standard-of-care radiographic imaging per oncology guidelines, evidence of residual disease before ACT, or death due to any cause.

The prespecified sample size of 80 study participants was intended to estimate the sensitivity and specificity of xM in predicting recurrence in stage II and III colorectal cancer from prespecified LMT and longitudinally. The sample size of 80 patients was obtained to allow the margin of error for the 95% confidence interval (CI) on sensitivity to be within 32% points under the assumption that the xM assay will provide a sensitivity of 57.5% and specificity of 95%. The study participants were randomly selected with enrichment for a recurrence rate of 50% compared with the 24% recurrence rate observed in GALAXY (20). Enrichment for recurrences was implemented to increase the precision in estimating sensitivity. Thus, the positive predictive value (PPV), negative predictive value (NPV), and HR were biased in a sample enriched for recurrences and these estimates were adjusted for the known lower recurrence rate. The sampling was stratified by stage and the stage II:III recurrent/nonrecurrent ratio to that observed in GALAXY was maintained. Specifically, 34% of recurrent study participants in GALAXY had stage II disease, whereas 66% had stage III. In the nonrecurrent cohort, 44% carried a stage II diagnosis, whereas 56% were diagnosed with stage III.

The evaluable population for longitudinal analysis was predefined, and the final MRD call was based on treatment (surgery-only vs. surgery and ACT), MRD status at LMT (when required for longitudinal analysis), and sample availability postoperatively for surgery-only study participants or after ACT for all others (Supplementary Table S1). For surgery-only study participants, final longitudinal MRD calls were based on all time points after surgery, up to and including the day of recurrence. MRD⁺ study participants were those who had at least one MRD⁺ result at any time point after definitive therapy (defined as surgery-only or surgery followed by ACT). MRD⁻ study participants were those with at least two evaluable samples and were persistently negative for MRD across all observed time points after definitive therapy. For study participants who received ACT, final longitudinal MRD calls were based on all time points after ACT (including the last day of ACT administration) and up to the time of clinical recurrence including the day recurrence was detected via radiographic imaging. MRD⁺ study participants were those who had an MRD⁺ result at any time point after ACT (including the last day of ACT), whereas MRD⁻ study participants were those with at least two evaluable samples and were persistently MRD⁻ across all observed time points after ACT (including the last day of ACT administration). The only exception is an ACT-treated study participant with an MRD⁻ LMT call, as predefined in our statistical analysis plan. In this case, only one MRD⁻ sample after ACT is sufficient to determine the final MRD⁻ status. Due to the variability of follow-up, samples collected within a window of up to 6 weeks after recurrence scan were utilized in determining final MRD status.

Sample collection/preparation

For each study participant, cell-free DNA (cfDNA) was extracted from approximately 8 mL of plasma banked for secondary studies of the GALAXY study and quantified to determine mass yields. For the methylation workflow (xM-methyl), 20 ng of mass input was prepared, and the remaining mass input (20–30 ng) was prepared and run on the genomic workflow (xM-variant; Fig. 1A).

Figure 1.

xM workflow and pipelines. A, xM dual workflow showing xM-methyl and xM-variant. The integration of these two workflows results in a binary MRD call. B, Schematic showing the different fragment level methyl configurations between regions that are similar in normal and tumor conditions and differentially methylated regions. C, Schematic explaining how the probability of a fragment being tumor derived is determined based on the observed number of CpG sites and the methylation status of those sites on a fragment, using representative training data from both tumor and normal samples. D, Workflow at longitudinal time points.

Genomic variant workflow: xM-variant assay

The xM-variant assay is a 0.3-Mbp panel that covers informative regions in greater than 100 genes. The assay uses New England Biolabs’ NEBNext Ultra II DNA Library Prep Kit to prepare cfDNA libraries that are sequenced on Illumina NovaSeq sequencers.

Methylation workflow: xM-methyl assay

The xM-methyl assay is a 6-Mbp panel based on the Twist Alliance Pan-cancer Methylation Panel and further customized with additional differentially methylated regions. The assay uses New England BioLabs’ NEBNext Enzymatic Methyl-Seq Kit to prepare cfDNA libraries that are sequenced on Illumina NovaSeq sequencers.

Model training

The xM-methyl model was trained on 86 tumor samples and 86 presumed-healthy or normal samples. To minimize the risk of confounding factors, tumor and normal samples were demographically matched on age, race, and gender. All samples were sequenced and processed by the assay and bioinformatics pipeline described below. For each sample and each region, fragment features such as the number of CpG covered and the number of CpG methylated were extracted (Fig. 1A and B). For each region and each sample class type (tumor and normal), a two-dimensional probability matrix was generated and a two-dimensional kernel density estimation (KDE) smoothing was applied (Fig. 1C). Tumor KDE and normal KDE were combined to create one probability distribution per region that represents the probability space of a fragment with a certain characteristic being tumor-derived versus normal-derived (Fig. 1A and C).

Methylation calling and model prediction pipeline

After sequencing, adapter-trimmed FASTQ files were aligned to hg38 using bwa-meth. PCR-duplicated reads were marked using SAMTOOLS (RRID: SCR_002105) to generate deduplicated BAMS that passed quality control (QC) criteria. A custom script to parse fragment-level information and MethylDackel was used to extract methylated calls, fragment-level, and sample-level features that were used for sample-level QC and as input into the model prediction pipeline. The model prediction pipeline including noise suppression algorithms was used to calculate the number of Unlikely Fragments Per Million.

xM-methyl MRD classification

The Unlikely Fragments Per Million threshold for calling a sample MRD⁺ was set to target an analytical specificity of ≥98% using a set of 126 training normal samples that were not used in model training or in establishing the ChG corrector.

Variant calling and model prediction pipeline

The bioinformatics pipeline was optimized and tuned to detect variants that indicate the presence of colorectal cancer with high specificity. If any candidate MRD variant is identified, the xM-variant MRD call is MRD⁺.

The process for identifying MRD variants started with predefining a set of genomic intervals where variants appear at high prevalence in colorectal cancer. A set of 8,674 training records, generated from prior solid tumor sequencing of colorectal cancer stage I to IV samples, were used to calculate pretest odds for each interval. Noisy intervals were removed based on a set of 40 normal training samples. In total, 287 genomic intervals ranging between 1 and 7 bp of variants with pretest odds >0.001 were selected as high-confidence colorectal cancer regions.

Clonal hematopoiesis (CH) variants were excluded as candidate variants if they had a historical detection frequency of >20% of being CH. The historical CH detection frequency was learned by leveraging Tempus' real-world multimodal database, which includes records of patients with solid tumors, cfDNA, and buffy coat sequencing.

Workflow utilized at LMT versus longitudinal samples

The dual workflow (methylation and genomic variant; Fig. 1A) was utilized by the xM assay for the LMT, which is the time of highest clinical actionability. For all subsequent time points, the xM assay used the methylation workflow alone (Fig. 1D).

Statistical analyses

For primary endpoint analysis, clinical performance (sensitivity, specificity, PPV, and NPV) was estimated with the corresponding 95% CI (exact 95% CI) at LMT using the evaluable population. For secondary endpoint analysis, Kaplan–Meier estimates were used to present DFS by MRD status at LMT in the evaluable population. DFS was measured from the surgery date to the date of first recurrence (via radiographic imaging) or death from any cause, whichever occurs first. Study participants were censored at the date of the last follow-up. The median follow-up for all 80 study participants was 22.1 months with a data cutoff point of 24 months. In addition, the HR between MRD⁺ and MRD⁻ (reference level) was calculated after adjusting for the anticipated true recurrence rate of 24%. MRD classification was generated for each sample without exposure to clinical outcomes to ensure unbiased analysis.

Definitions for landmark and longitudinal true-positive, false-positive, true-negative, and false-negative calls can be found in Supplementary Table S2.

Data availability

The human sequence data used in this study (including deidentified data) are not publicly available due to patient privacy and proprietary requirements. When possible, derived data supporting the findings of this study have been made available within the article and its Supplementary figures/tables. Further details can be made available through the corresponding author upon reasonable request.

Results

Of the original 80 study participants in the cohort, 69 were evaluable at LMT and 74 were evaluable longitudinally. Eleven study participants from the biobanked samples were excluded due to an invalid assay result including one LMT sample that did not have sufficient cfDNA to sequence the genomic variant component of the assay and 10 LMT samples that did not pass QC.

Overall, the study participants had a mean age of 67.7 years (range, 37–89) at enrollment with 29 females (36.3%) and 51 males (63.8%) (Table 1). In terms of stage distribution, 32 (40.0%) of study participants were stage II and 48 (60.0%) were stage III. Most of the study participants [n = 72 (90.0%)] had a microsatellite instability stable status with the remaining [n = 8 (10%)] resulting as MSI-high. Additionally, 37 study participants (46.3%) had an RAS-positive mutation status. The Pearson correlation coefficient between RAS and ctDNA was 0.213 suggesting weak correlation (P = 0.08). Eight study participants (10.0%) had BRAF-positive status. All but one study participant had a histologic diagnosis of adenocarcinoma (98.9%); a single study participant had a diagnosis of adenosquamous carcinoma. Additionally, 32 (40.0%) received ACT.

We sequenced the available longitudinal samples for all 80 study participants, of which 74 were evaluable. For the longitudinal analysis, we evaluated three endpoints. We analyzed (i) the longitudinal clinical performance of the Tempus xM assay in stage II and III colorectal cancer, (ii) the lead time or median time from first positive MRD result to clinical recurrence detected via radiographic imaging per oncology guidelines, and (iii) DFS by CEA status at 12 weeks (abnormal: CEA >5.0 ng/mL, normal: CEA ≤5.0 ng/mL) to DFS by MRD status at 12 weeks.

Sensitivity and specificity analysis

Primary LMT analysis

Out of the 36 evaluable study participants with known recurrence, 22 were MRD⁺ at the LMT, and out of the 33 nonrecurrent study participants, 29 were characterized as MRD⁻ at LMT (sensitivity of 61.1% and specificity of 87.9%, Fig. 2; Supplementary Table S3). At LMT, the PPV and NPV adjusted by the anticipated recurrence rate of 24% were 0.61 and 0.88, respectively. For stage II study participants specifically, the adjusted PPV was 75.3% with an adjusted NPV of 89.2%. For stage III study participants, the adjusted PPV was 52.8% with an adjusted NPV of 86.6% (Supplementary Table S3). Clinical sensitivity and specificity by cancer stage are shown in Supplementary Fig. S1.

Figure 2.

Clinical sensitivity and specificity in recurrent and nonrecurrent study participants. A, Clinical sensitivity of xM in recurrent CRC study participants at LMT (left) and longitudinal time points (right). Arrows show the number of study participants who switched MRD status between LMT and longitudinal time points. LMT clinical sensitivity was 61.1% (22/36 study participants); longitudinal clinical sensitivity was 83.3% (30/36 study participants). B, Clinical specificity of xM in nonrecurrent CRC study participants at LMT (left) and longitudinal time points (right). Arrows show the number of study participants who switched MRD status between LMT and longitudinal time points. LMT clinical specificity was 87.9% (29/33 study participants); longitudinal specificity was 89.5% (34/38 study participants). CRC, colorectal cancer.

Longitudinal analysis

A total of six study participants were not evaluable for longitudinal analysis. Three of those study participants had undergone surgery alone and the other three had received ACT after resection. Of the study participants who only had undergone surgery, two were excluded because they had no available samples after LMT and the MRD status was negative at LMT. The third study participant was excluded because their only sample available past the MRD⁻ result at LMT was invalid. From the ACT group, one study participant was excluded because their LMT sample was invalid, and two additional study participants were excluded because they did not have valid samples available after completion of ACT. From the 74 study participants evaluable for longitudinal analysis, there was a range of one to seven available samples per study participant not counting the presurgery baseline, with a mean of 4.5 evaluable samples per study participant.

MRD longitudinal sensitivity (defined in Supplementary Table S1) was 83.3% (95% CI, 67.2%–93.6%), in which 30 out of 36 evaluable recurrent study participants were classified as MRD⁺ (Fig. 2A). Of the 14 false-negative results assessed from the LMT-only analysis, seven study participants had a true positive longitudinal finding, three remained false negatives, and four were not evaluable for longitudinal analysis (Fig. 2A). The longitudinal sensitivity did differ by stage with stage II study participants having a longitudinal sensitivity of 91.7% (95% CI, 61.5%–99.8%) and stage III study participants experiencing a longitudinal sensitivity of detecting MRD in ctDNA of 79.2% (95% CI, 57.8%–92.9%; Supplementary Table S3).

MRD longitudinal specificity was 89.5% (95% CI, 75.2%–97.1%), as out of the 38 nonrecurrent study participants, 34 were MRD⁻ based on the final longitudinal MRD call (Fig. 2B). Longitudinal specificity also varied by stage with specificity of 88.2% (95% CI, 63.6%–98.5%) for stage II study participants and 90.5% for stage III study participants (95% CI, 69.6%–98.8%; Supplementary Table S3). Longitudinal adjusted PPV was 71.4% and adjusted NPV was 94.4% overall (71.1% and 97.1% for stage II and 72.4% and 93.2% for stage III, respectively).

Recurrence analysis

Lead time from first MRD⁺ result to clinical recurrence

Lead time measures the follow-up time from the time of the first positive MRD result to clinical recurrence detected by radiographic imaging per oncology guidelines among recurrent study participants with a longitudinal MRD⁺ result. For study participants who received ACT, lead time is calculated from the time of the first positive MRD after the end of ACT. Thus, lead time was calculated on the 30 longitudinal MRD⁺ study participants among the 41 recurrences (Fig. 3A). The overall median lead time, based on 30 participants, was 4.77 months. The median lead time for the 22 study participants who underwent surgery only was 5.30 and 0.01 months for the eight study participants who received ACT. A swimmer plot of all clinically recurrent study participants is shown in Fig. 3B. Nonrecurrent study participants are shown in the swimmer plot in Supplementary Fig. S2.

Figure 3.

Clinical recurrence. A, Distribution of lead time (time from the first MRD⁺ call to date of recurrence or death) for true-positive (TP, n = 30) study participants. Overall median lead time defined from the first MRD⁺ to recurrence is 4.77 months. For study participants with surgery-only treatment, the median lead time is 5.30 months. B, Swimmer plot shows xM MRD assay results, recurrence status, and timing of ACT for all recurrent study participants. The blue bars to the left of the swimmer plot highlight patients in each MRD call category who received ACT. C, Graph shows the most common sites of recurrence in study participants with recurrent CRC based on the MRD call at LMT. Although the liver was the most common site of recurrence, there is no statistical difference between the liver and the other locations combined (liver vs. ovary, peritoneum + lung: P = 0.07; liver vs. ovary, peritoneum: P = 0.11; liver vs. lung: P = 0.18). CRC, colorectal cancer.

Site of recurrence according to MRD status

We reviewed the most common metastatic recurrence sites to determine site-specific ctDNA detection. The sites that we included in the analysis were liver, ovary/peritoneum, lung, and local or lymph node-only recurrences. Out of the 15 liver-only recurrences at LMT, 11 were MRD⁺ and four were MRD⁻. Thus, the true-positive rate is 73.3% (11/15) for liver-only recurrences (Fig. 3C). Out of the eight ovarian or peritoneal metastatic recurrences, three were MRD⁺ at LMT and five were MRD⁻. Out of the seven lung-only metastatic recurrences, three were MRD⁺ at LMT and four were MRD⁻. The combined true-positive rate at LMT for ovarian, peritoneal, and lung metastatic sites was 40% (6/15). The true- positive rates between these two major groups (liver only vs. lung, ovary, or peritoneum) were compared using Fisher’s exact test. However, the results did not reach statistical significance (P = 0.07). Longitudinally, true-positive rates are higher across all recurrence sites compared with LMT true-positive rates, but the overall patterns are unchanged (Supplementary Fig. S3).

Survival analysis

We analyzed the association between xM MRD detection status and DFS and determined that MRD⁻ status correlated with longer DFS with an adjusted HR of 7.28 (12.25 for stage II and 5.21 for stage III study participants; Supplementary Table S3). The HR was adjusted to account for the 24% recurrence rate observed in the study. The longitudinal adjusted HR was 21.9 overall (42.24 for stage II study participants and 18.35 for stage III study participants).

Association between 12-week CEA and DFS (a post hoc analysis)

CEA is a known disease biomarker for study participants with colorectal cancer. For this analysis, we focused on the 66 study participants who had an MRD call available at 12 weeks postsurgical resection. Of those study participants, four had missing CEA values at 12 weeks postsurgical resection, and one experienced recurrence before 12 weeks (recurrence took place at 8 weeks) leaving 61 study participants with CEA data to analyze. Abnormal CEA results were classified as those with values >5.0 ng/mL, whereas normal CEA results were those </= 5.0 ng/mL.

CEA normal status at 12 weeks postsurgery was associated with longer DFS than CEA abnormal status (adjusted HR, 2.13; Fig. 4A). MRD⁻ study participants at 12 weeks postsurgery had longer DFS than MRD⁺ study participants at the same time point (adjusted HR, 9.69; Fig. 4B). The adjusted HR for MRD relative to CEA at 12 weeks postsurgery was >4.5-fold. The adjusted median survival for MRD⁺ study participants was 6.3 months (25.1 weeks) and not reached within the 72 weeks (18 months) follow-up for MRD⁻ study participants at 12 weeks postsurgery.

Figure 4.

DFS and xM comparison to CEA. A, The adjusted median DFS time for MRD⁺ study participants is 25.1 weeks (6.3 months) vs. not reached within 72 weeks (18 months) for MRD⁻ study participants. Shadowing surrounding each curve represents 95% CI for Kaplan–Meier estimates. B, Adjusted HR for xM MRD is nearly fivefold higher than CEA testing at 12 weeks postsurgery. Adjusted HR* is the HR adjusted by anticipated true recurrence rate (24%). Shadowing surrounding each curve represents 95% CI for Kaplan–Meier estimates.

Preoperative MRD association with DFS

Out of the 67 available presurgery samples, 57 (85.1%) were MRD⁺. Adjusted HR for MRD status versus DFS was 0.82. Presurgery MRD status does not seem to be a prognostic indicator of recurrence (Supplementary Table S4).

Discussion

Relative to tumor-informed MRD assays, tumor-naïve ctDNA-based MRD assays promise faster turnaround time, require no tissue acquisition, and better capture overall tumor heterogeneity. Despite these advantages, tumor-naïve MRD assays have been viewed as inferior, largely due to their lower analytical sensitivity relative to tumor-informed MRD assays and the general assumption that the differences observed in analytical performance also translate directly to differences in clinical performance. However, there is no scientific consensus about whether a lower limit of detection in different types of solid tumor MRD assays, which measure fundamentally different biological analytes, correlates to inferior clinical performance, nor which analytical metric could be most informative clinically. Despite the heterogeneity in the lower limit of detection among different types of MRD assays, there is a range of acceptable clinical performance that can be reliably achieved even with these analytical differences (30, 31).

No singular MRD assay can satisfy all clinical use cases. Various considerations such as tumor biology heterogeneity, ctDNA shedding rate, tissue quality, and scarcity all need to be accounted for when selecting an optimal MRD assay for specific populations of patients with cancer. Moreover, optimizing performance, QC, and clinical utilization of MRD assays is a significant undertaking, considering the complexity of tissue acquisition and the standardization of insurance coverage for next-generation sequencing assays within a global health care system. As MRD assays move into the companion diagnostic setting for novel therapeutics in early-stage solid cancer, technical ease and standardization of MRD assays will become an important consideration in large phase III studies and global drug approvals.

By leveraging the strengths of a tumor-naïve approach, we have demonstrated that the xM assay provided robust clinical performance, with a clinical sensitivity of 61.1% to detect recurrence at LMT, increasing to 83.3% when examined longitudinally. Utilizing xM at LMT, a time of high clinical actionability, is a critical moment between physicians and their patients when deciding on adjuvant treatment based on standard clinicopathologic criteria, including stage and additional risk factors (e.g., perforation, gastrointestinal obstruction, and other criteria; ref. 32). Implementing MRD status results from a rapid, high-performance assay at LMT enables clinicians to tailor treatment options, which may improve patient outcomes. In addition, at LMT, genomic variant analysis from this assay in MRD⁺ patients may provide valuable molecular information that could serve as a prognostic or predictive biomarker and may impact future treatment management. As more data from randomized ctDNA-guided clinical trials become available, xM MRD status is uniquely positioned as a biomarker to be included in trial protocols.

It is critical to follow patients with early-stage colorectal cancer longitudinally as part of routine surveillance for a minimum of 2 years and up to 5 years as recommended by NCCN Guidelines (32). We have shown a robust longitudinal clinical sensitivity of 83.3% within the standard surveillance window, capturing the majority of patients who relapsed ahead of radiographic evidence of recurrence per oncology guidelines. Following MRD status longitudinally helps identify MRD⁺ patients earlier who may require escalation of therapy and/or increased radiographic surveillance frequency in addition to standard-of-care management. As important, patients who remain persistently MRD⁻ longitudinally may not need additional therapy and clinical observation may be sufficient. ctDNA-guided management represents a shift in oncology care from overtreating to biomarker-directed therapy that may spare acute and long-term chemotherapy-related toxicity, including platinum neurotoxicity and related health care costs.

With a median lead time of 4.8 months from the first ctDNA positive result to clinical recurrence, MRD testing with the xM assay may present oncologists with a window of opportunity to escalate care and optimize treatment postoperatively. However, we also demonstrated that a change in MRD status from negative to positive, or persistent MRD positivity from xM testing, is a prognostic biomarker for patients who are likely to experience imminent clinical recurrence. In addition, we have shown that MRD status from xM has a fourfold higher HR in predicting DFS than longitudinal CEA level at 12 weeks after curative intent surgical resection. Future research may examine correlation at additional time points after 12 weeks, as guidelines recommend CEA testing every 3 months for a minimum of 2 years.

Conclusions

The novel tumor-naïve MRD assay, Tempus xM, is a rapid, finely tuned MRD test that accurately detected ctDNA in study participants with clinical recurrence at LMT with a clinical sensitivity of 61.1%. Additionally, xM demonstrated a clinical sensitivity of 83.3% and specificity of 89.5% when used longitudinally. There was a superior correlation of MRD status to DFS compared with standard-of-care CEA level. We are planning a larger clinical validation study incorporating additional samples and data from CIRCULATE-Japan’s GALAXY, designed to enhance statistical robustness. This will enable a more comprehensive evaluation of the relationship between ACT and DFS benefit in MRD⁺ versus MRD⁻ colorectal cancer participants.

xM’s MRD model has been trained on a continuously evolving real-world multimodal database enabling accurate corrections for specific colorectal cancer–related confounders including germline variants and CH of indeterminate potential, contributing to the assay’s robust performance despite biological artifacts. As Tempus’ multimodal database continues to grow, we may be able to further optimize the assay performance in the future.

Supplementary Material

Supplementary Figure S1

Figure S1. Clinical specificity and sensitivity by cancer stage at LMT and longitudinal time points

Supplementary Figure S2

Figure S2. MRD status over time in non-recurrent study participants

Supplementary Figure S3

Figure S3. Site of clinical recurrence in longitudinal study

Supplementary Table S1

Table S1. MRD call definitions

Supplementary Table S2

Table S2. Definitions for landmark and longitudinal true positive, false positive, true negative, and false negative cases

Supplementary Table S3

Table S3. Summary statistics at LMT and longitudinal time points

Supplementary Table S4

Table S4. Pre-surgery MRD association to DFS

Authors’ Disclosures

Y. Nakamura reports grants from Tempus AI, Inc. during the conduct of the study, as well as personal fees from Guardant Health Pte Ltd., Natera, Inc., Roche Ltd., Seagen, Inc., Premo Partners, Inc., Takeda Pharmaceutical Co., Ltd., Exact Sciences Corporation, Gilead Sciences, Inc., MSD K.K., and Eisai Co., Ltd.; personal fees from Zeria Pharmaceutical Co., Ltd., Miyarisan Pharmaceutical Co., Ltd., Merck Biopharma Co., Ltd., CareNet, Inc., Hisamitsu Pharmaceutical Co., Inc., Taiho Pharmaceutical Co., Ltd., Becton, Dickinson and Company, and Guardant Health Japan Corp.; personal fees and nonfinancial support from Daiichi Sankyo Co., Ltd., and Chugai Pharmaceutical Co., Ltd.; and grants from Seagen, Inc., Genomedia Inc., Guardant Health AMEA, Inc., Guardant Health, Inc., and Roche Diagnostics K.K. outside the submitted work. K. Kaneva reports personal fees and other support from Tempus AI, Inc.; other support from Tempus AI, Inc. during the conduct of the study; personal fees and other support from Tempus AI, Inc.; and other support from Tempus AI, Inc. outside the submitted work. C. Lo reports personal fees and other support from Tempus AI, Inc., other support from Tempus AI, Inc. outside the submitted work, and a patent for 18/913,883 pending. D. Neems reports personal fees and other support from Tempus AI, Inc., other support from Tempus AI, Inc. outside the submitted work, and a patent for 18/913,883 pending. J.E. Freaney reports personal fees and other support from Tempus AI, Inc. and other support from Tempus AI, Inc. outside the submitted work. S.W. Hyun reports personal fees from Tempus AI, Inc. and Tempus AI, Inc. outside the submitted work. F. Islam reports personal fees from Tempus AI, Inc. and other support from Tempus AI, Inc. outside the submitted work. J. Yamada-Hanff reports personal fees and other support from Tempus AI, Inc., other support from Tempus AI, Inc. during the conduct of the study, and a patent for 18/913,883 pending. T.M. Driessen reports personal fees from Tempus AI, Inc., other support from Tempus AI, Inc. outside the submitted work, and a patent for 18/913,883 pending. A. Sonnenschein reports personal fees from Tempus AI, Inc., other support from Tempus AI, Inc. outside the submitted work, and a patent for 18/913,883 pending. D.F. DeSantis reports personal fees from Tempus AI, Inc. and other support from Tempus AI, Inc. outside the submitted work. J. Watanabe reports grants and personal fees from Medtronic, personal fees from Johnson and Johnson, Eli Lilly and Company, and Takeda, and grants from AMCO, TERUMO, and Stryker Japan outside the submitted work. M. Kotaka reports personal fees from Chugai and Eli Lilly and Company outside the submitted work. S. Mishima reports personal fees from Taiho Pharmaceutical, Chugai Pharmaceutical, and Eli Lilly and Company outside the submitted work. H. Bando reports personal fees from Ono Pharmaceutical, Taiho Pharmaceutical, and Eli Lilly Japan outside the submitted work. K. Yamazaki reports personal fees from Chugai Pharma, Takeda, Yakult, Taiho, Daiichi Sankyo, Merck Biopharma, Eli Lilly and Company, Ono Pharmaceutical, MSD, and Bristol Myers and Squibb outside the submitted work. H. Taniguchi reports personal fees from Chugai Pharmaceutical, Ono Pharmaceutical, and Eli Lilly and Company; grants and personal fees from Takeda; and grants from Daiichi Sankyo outside the submitted work. T. Kato reports other support from Chugai Pharmaceutical Co., Ltd, Takeda Pharmaceutical Company Limited, Ono Pharmaceutical Co., Eli Lilly and Company, and Taiho Pharmaceutical outside the submitted work. C. Sangli reports personal fees from Tempus AI outside the submitted work. R. Tell reports personal fees and other support from Tempus AI, other support from Tempus AI outside the submitted work, and a patent for 18/913,883 pending. R. Blidner reports personal fees from Tempus during the conduct of the study and personal fees from Tempus outside the submitted work. T. Yoshino reports grants from Amgen K.K., Bristol Myers Squibb K.K., Daiichi Sankyo Co., Ltd., Eisai Co., Ltd., FALCO Biosystems Ltd., Genomedia Inc., Medical & Biological Laboratories, Co., Ltd., Merus N.V., Molecular Health GmbH, Nippon Boehringer Ingelheim Co., Ltd., Pfizer Japan Inc., Roche Diagnostics K.K., Sanofi K.K., Sysmex Corp., Taiho Pharmaceutical Co., Ltd.; grants and personal fees from Chugai Pharmaceutical Co., Ltd., MSD K.K., Takeda Pharmaceutical Co., Ltd., Ono Pharmaceutical, and Co., Ltd.; and personal fees from Merck Biopharma Co., Ltd., and Bayer Yakuhin, Ltd. outside the submitted work. K. Sasser reports other support from Tempus outside the submitted work. E. Oki reports other support from Chugai Pharmaceutical, Ono Pharmaceutical, Eli Lilly and Company, Bristol Myers Squibb, and Takeda Pharmaceutical outside the submitted work. H. Nimeiri reports other support from Tempus AI, AbbVie, and Northwestern Medicine outside the submitted work. No disclosures were reported by the other authors.

Authors’ Contributions

Y. Nakamura: Resources, data curation, visualization, writing–original draft, project administration, writing–review and editing. K. Kaneva: Conceptualization, data curation, formal analysis, supervision, validation, investigation, methodology, writing–original draft, project administration, writing–review and editing. C. Lo: Resources, data curation, software, formal analysis, supervision, visualization, writing–original draft, project administration, writing–review and editing. D. Neems: Resources, data curation, software, formal analysis, visualization, methodology, writing–original draft, project administration, writing–review and editing. J.E. Freaney: Conceptualization, resources, data curation, supervision, funding acquisition, visualization, writing–original draft, project administration, writing–review and editing. H. Boulos: Resources, supervision, investigation, methodology, writing–review and editing. S.W. Hyun: Formal analysis, supervision, methodology, writing–review and editing. F. Islam: Resources, data curation, formal analysis, visualization, writing–original draft, project administration, writing–review and editing. J. Yamada-Hanff: Resources, data curation, software, formal analysis, investigation, visualization, writing–original draft, project administration, writing–review and editing. T.M. Driessen: Resources, data curation, software, formal analysis, visualization, writing–original draft, project administration, writing–review and editing. A. Sonnenschein: Software, formal analysis, writing–review and editing. D.F. DeSantis: Resources, data curation, visualization, writing–original draft, project administration, writing–review and editing. D. Kotani: Resources, data curation, writing–original draft, writing–review and editing. J. Watanabe: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. M. Kotaka: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. S. Mishima: Conceptualization, resources, data curation, formal analysis, supervision, validation, investigation, methodology, writing–original draft, project administration, writing–review and editing. H. Bando: Conceptualization, resources, data curation, software, formal analysis, supervision, validation, investigation, methodology, writing–original draft, project administration, writing–review and editing. K. Yamazaki: Conceptualization, resources, data curation, software, formal analysis, supervision, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. H. Taniguchi: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. I. Takemasa: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. T. Kato: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. C. Sangli: Conceptualization, supervision, methodology, writing–review and editing. R. Tell: Resources, supervision, writing–review and editing. R. Blidner: Supervision, project administration, writing–review and editing. T. Yoshino: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. K. Sasser: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. E. Oki: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. H. Nimeiri: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing.