Next-generation Multi-target Stool DNA Panel Accurately Detects Colorectal Cancer and Advanced Precancerous Lesions

Colorectal cancer remains the second most common cause of cancer-related deaths for men and women in the United States, with an estimated 153,020 new colorectal cancer diagnoses and 52,550 colorectal cancer–related deaths occurring in 2023 (1). National guidelines recommend that adults at average risk ages 45 years or older receive colorectal cancer screening either through direct visual examinations (e.g., colonoscopy) or noninvasive, stool-based tests (2, 3). Population-based colorectal cancer screening can reduce mortality through the detection and removal of advanced precancerous lesions (APL) and early-stage cancer (4, 5). However, colorectal cancer screening rates remain below the nationwide goal established by the National Colorectal Cancer Roundtable (4). Offering noninvasive testing options, such as fecal immunochemical testing (FIT), has been shown to increase screening participation (6) although opportunity for improvement exists with respect to FIT-based detection of APLs and patient adherence with annual screening (7–14), as typically recommended by national guidelines (2, 4).

The multi-target stool DNA test (mt-sDNA; commercialized as Cologuard, Exact Sciences) is a noninvasive option for average-risk colorectal cancer screening that is included in recommendations from the American Cancer Society, U.S. Preventive Services Task Force, and other major organizations (2, 3). Modeling studies and major guidelines support use of the mt-sDNA test every 3 years (2, 3, 15), and patient adherence with mt-sDNA testing can exceed 70% (16, 17).

Since receiving approval from the FDA in 2014 (18), more than 10 million people have been screened using the mt-sDNA test (19, 20). The mt-sDNA test combines quantification of the methylated DNA markers (MDM) NDRG4 and BMP3, DNA mutations in KRAS, ACTB (a marker of human DNA input), and fecal hemoglobin (Hb) into an algorithm to generate a qualitative test result of positive or negative (12). In the pivotal trial establishing test performance (“DeeP-C”, NCT01397747), the first-generation mt-sDNA test demonstrated significantly higher sensitivity for colorectal cancer and APLs than the FIT (OC FIT-CHEK, Polymedco; 92% vs. 74% for colorectal cancer, P = 0.002; 42% vs. 24% for APLs, P < 0.001; ref. 12). The sensitivity of first-generation mt-sDNA for APLs increases with size (42%–50% for APLs ≥1 cm in size, 63%–67% for >2 cm, and 70%–80% for >3 cm; refs. 12, 13, 21). These performance estimates have been replicated in subsequent studies (13, 14).

Since DeeP-C, novel MDMs have been discovered using next-generation sequencing of DNA extracted from non-neoplastic, primary colorectal cancer, and APL tissues (22). These biomarkers have appeared broadly informative in tissue studies and an early-phase stool pilot study (23). Here, we estimated the performance of a next-generation mt-sDNA panel incorporating novel MDMs, normalized to methylated ZDHHC1 (23), in combination with Hb, using samples from two large studies (Fig. 1). The first, a blinded case–control experiment in archival specimens, was a marker selection study (MSS) of several highly discriminant candidate MDMs to include in the panel. The second, a follow-up independent verification study (VS), emphasized panel performance in blinded archival colorectal cancer specimens weighted to early-stage and prospectively collected APLs, simulating a screening population.

The MSS included archived negative (non-APL, non-neoplastic) stool samples that were prospectively collected as part of the DeeP-C study (12). Additional archival stool samples from case patients with colorectal cancer or APL were archived as part of previously approved protocols (NCT01260168 and NCT02503631). The marker panel collection study (NCT01260168) was conducted according to Good Clinical Practice (GCP) Guidelines, the Declaration of Helsinki, and US 21 Code of Federal Regulations. The stool sample collection study (NCT02503631) was conducted in accordance with applicable legal and regulatory requirements, as well as the general principles set forth in the International Ethical Guidelines for Biomedical Research Involving Human Subjects, GCP guidelines, and the Declaration of Helsinki. All collection sites had Institutional Review Board approval and all eligible subjects provided written informed consent to participate in the studies. Samples were required to have a previous Hb result or sufficient remaining volume for Hb testing, historical results for first-generation mt-sDNA methylation and mutation markers, and known staging, architecture, and size information for colorectal cancer and APLs. Samples were selected to generate a uniform age distribution of participants to better estimate any potential association of DNA methylation with age. Sample selection across cases and controls was designed to include at least 50% from non-White participants. Samples were blinded and randomized prior to testing.

The VS included archived colorectal cancer stool samples which were also archived under the previously approved protocols (NCT01260168 and NCT02503631). All other VS samples (APL, non-APL, non-neoplastic) were prospectively collected as part of the DeeP-C study (12). Samples were required to have either a previous Hb result or sufficient volume for Hb testing. Samples were selected such that the sex, race/ethnicity, and age distributions of participants aligned with those of the DeeP-C study. In addition, the distribution of lesion size and the relative frequencies of pathology subcategories for colorectal cancer, APLs, and control samples were designed to match those of DeeP-C (12).

Diagnoses of case and control endpoints for both studies were determined by colonoscopy and histopathologic review by gastrointestinal pathologists at the participating sites. Colorectal cancer was classified according to the 8th Edition of the American Joint Committee on Cancer staging criteria (24). APLs were subdivided into four categories, with hierarchical categorization applied when indicated: (i) high-grade dysplasia (including carcinoma in situ), any size (“HGD”); (ii) villous growth pattern (≥25%), any size (“villous”); (iii) ≥1.0 cm in size (“tubular”); and (iv) sessile serrated lesion, ≥1.0 cm in size (“SSL”). Non-APLs were subdivided into “1 or 2 adenomas, >0.5 cm or <1.0 cm in size,” “≥3 adenomas, <1.0 cm in size,” and “1 or 2 adenomas, ≤0.5 cm in size.” Non-neoplasia samples were classified as “No neoplasia upon histopathologic review (hyperplasia)” or “No findings in colonoscopy; no histopathologic review”. Specificity was calculated either for the absence of advanced neoplasia (non-APL/non-neoplasia/no findings in colonoscopy), or for no neoplastic findings (non-neoplasia/no findings in colonoscopy). Colorectal cancer and APL results were confirmed centrally, with diagnostic disagreements resolved by a consensus of at least two central pathologists.

Samples for both the MSS and VS were processed similarly to quantify MDMs and protein markers (Supplementary Fig. S1). Briefly, whole stool samples were homogenized, and inhibitors were removed that would impact downstream DNA processing. Next, targeted capture of MDMs was performed. After bisulfite conversion and purification, MDMs were quantified using long‐probe quantitative amplified signal (LQAS) assays, or the quantitative allele-specific real-time target and signal (QuARTS) amplification assays (25). Fecal Hb was collected and quantified using the first-generation mt-sDNA Hb collection device, buffer, and ELISA as described previously (12).

The objective of the MSS was to identify the best set of markers to distinguish between colorectal cancer/APL case and control samples. The study included approximately 20 colorectal cancer/APL cases for each MDM marker.

Fifty-two markers were assessed in preclinical experiments on the basis of AUC rank, signal to noise ratio (>50×), and control sample background methylation of < 1% (22). From these, 12 candidates were selected on the basis of marker discrimination, complementarity, signal to noise ratio, control sample background methylation levels, and rate of increase in background methylation with age, as was previously described for the first-generation mt-sDNA test (26).

Two approaches were evaluated. First, logistic regression was used to model colorectal cancer/APL (Yes/No) on all combinations of 12 candidate MDMs (BMP3, CHST2, CHST10, DMRTA2, LASS4, LRRC4, NDRG4, PDGFD, PPP2R25C, SDC2, SFMBT2, VAV3), the reference marker ZDHHC1, and Hb. The search was restricted to models that incorporated four markers or fewer. These models were then ranked according to their Akaike information criteria (AIC; lower AICs indicated a better fit to the data), which standardizes the likelihood for the number of regressors in the model.

The second approach used random forest modeling to estimate individual marker importance for predicting colorectal cancer/APL. Briefly, 500 bootstrap random samples of the data were used to train the random forest model with the 12 candidate markers. Once trained, the mean decrease in accuracy was estimated for an individual marker by randomly permuting its values while the remaining markers remained unchanged. The mean difference was calculated between permuted and unchanged prediction accuracy across the 500 bootstrap samples (27). Markers with a higher mean decrease in accuracy were considered more informative than markers with lower mean decrease in accuracy.

The VS was conducted to better estimate the sensitivity of the next-generation mt-sDNA panel. The available cases for the VS included 98 APL samples, which would provide a mean 95% confidence interval (CI) range for sensitivity of ±10%, 112 colorectal cancer cases (mean 95% CI sensitivity range ±5%), and 567 control cases (mean 95% CI specificity range of ±2.5%).

A logistic model with two explanatory variables was fitted to the VS data in which colorectal cancer and APL samples were considered positive, and all other samples were negative. Three fitted regression coefficients were used: the intercept, the contribution from Hb, and the contribution from averaged methylated gene markers (aMDM; defined as the arithmetic average of the three markers: LASS4, LRRC4, PPP2R5C). β-actin (ACTB), used as the reference marker for the first generation mt-sDNA test, was also evaluated to provide a comparison with the novel reference marker ZDHHC1. The final logistic model was cross-validated 4-fold, with stratified 75%/25% training/test set splits repeated 1,000 times.

The statistical analysis was performed using R statistical software. The generalized linear model function was used to fit the logistic classifiers. Other modeling software packages were used, including random forest, least absolute shrinkage and selection operator (LASSO), and neural networks, for alternative modeling approaches.

A limited deidentified dataset may be made available upon request, subject to approval of the study sponsor.

An overview of the development of the next-generation mt-sDNA panel is shown in Fig. 1. These included independent samples for the MSS to identify promising MDMs and the VS to estimate the performance of the final marker panel, with an emphasis on APLs.

The MSS included 232 case samples (120 colorectal cancers and 112 APLs) and 490 control samples (163 non-APLs and 327 negative; Supplementary Table S1). A table of participant characteristics is available in Supplementary Table S2. The boot-strapped logistic regression analysis of the MSS results identified three MDMs for the VS (LASS4, LRRC4, PPP2R5C, and reference marker ZDHHC1), as well as Hb. In addition, these three methylation markers exhibited some of the lowest increases in methylation with age. The logistic classification model showed a sensitivity of 91.7% (95% CI, 85.4–95.4) for colorectal cancer and 62.5% (53.3–70.9) for APLs, with 90.4% specificity for the absence of advanced neoplasia (87.5–92.7) in samples from patients with non-APLs or non-neoplastic findings; the specificity was 90.8% for non-neoplastic findings only (Supplementary Table S3). The AUC of the MSS ROC curves were 0.97 (0.95–0.99) for colorectal cancer and 0.82 (0.77–0.87) for APLs (Supplementary Fig. S2). Colorectal cancer sensitivity was 88.5% for stage I (n = 52), and 91.9% for stage II (n = 38). APL sensitivity was relatively uniform across subtypes; 52.9% for HGD, 66.7% for villous, 59.3% for tubular, and 64.3% for SSLs (Supplementary Fig. S3A and S3B). These performance values were nominally higher than those with ACTB as a reference marker, which showed sensitivities of 92.9% for colorectal cancer and 50.0% for APLs, with 90.2% specificity for non-advanced findings.

The VS included 112 colorectal cancers (49 stage I and 38 stage II), 98 APLs, 176 non-APLs, and 391 non-neoplastic samples (Table 1). The patient characteristics for the VS are available in Table 2. The average age (and SD) of patients was 65.5 ± 8.5 years for cases (n = 210) and 63.2 ± 8.7 years for controls; 57% of cases were male compared with 47% of controls. There was a similar distribution of racial and ethnic participants between case and control groups (Table 2).

All prespecified models (including random forest, LASSO, and neural networks) generated similar results; the generalized linear model was chosen to be presented here as it was the simplest model, and least likely to overestimate assay performance due to overfitting. The mean sensitivity from cross-validation of the next-generation mt-sDNA panel in the VS was 95.2% (89.5–97.9) for colorectal cancer and 57.2% (47.3–66.5) for APLs, with 89.8% (87.0–92.0) specificity in samples from patients with non-advanced findings (Table 3). The specificity in samples with non-neoplastic findings alone was 92.4% (Table 3). The AUCs of the VS ROC curves were 0.97 (0.95–0.99) for colorectal cancer and 0.80 (0.75–0.86) for APLs (Fig. 2). The next-generation mt-sDNA panel demonstrated high sensitivity for early-stage colorectal cancer, with 93.9% for stage I (n = 49) and 94.7% for stage II (n = 38; Fig. 3A). APL sensitivity by subtype was 83.3% for HGD, 70.6% for villous, 51.1% for tubular, and 30.8% for SSLs (Fig. 3B). Using ZDHHC1 as a reference marker produced nominally higher performance values than using ACTB as a reference marker, which showed sensitivities of 92.9% for colorectal cancer and 50.0% for APLs, with 90.1% specificity for the absence of advanced neoplasia.

Data from two independent archives of stool samples demonstrated that the next-generation mt-sDNA panel of MDM and protein markers showed high accuracy for discriminating colorectal cancer and APLs from non-neoplastic controls. Cross-validation analyses were utilized to minimize overfitting in the model, and the performance of the panel was assessed in an independent sample set that included a high proportion of early-stage colorectal cancer and APL cases to simulate an average-risk colorectal cancer screening population and minimize selection bias.

We observed a slightly lower APL sensitivity and slightly higher colorectal cancer sensitivity in the VS compared with the MSS. This discrepancy may be due in part to differences in lesion size, which was shown to correlate with test sensitivity in the DeeP-C study utilizing the first-generation mt-sDNA test (12). Lesion size was not controlled for in the MSS, which resulted in smaller colorectal cancer samples and larger APL samples than those included in the VS (Supplementary Table S3). Samples were selected to mimic the distribution of lesion sizes that would be expected in a screening population for the VS (Table 3; ref. 12). As a result, the VS likely provides a better estimate of colorectal cancer and APL sensitivity for the final next-generation mt-sDNA panel compared with the MSS.

Because the performance characteristics of the next-generation mt-sDNA panel have not been established in a screening cohort design study, it is premature to compare sensitivity and specificity results with those from the clinically available first-generation mt-sDNA test. However, anticipated differences between the first-generation mt-sDNA test and the next-generation mt-sDNA panel do warrant a brief discussion. First, the DNA marker reference gene in the next-generation mt-sDNA test has been changed—the currently approved first-generation mt-sDNA test uses ACTB while the next-generation mt-sDNA panel uses ZDHHC1. The ZDHHC1 marker is constitutively methylated in both normal colonic epithelia and colorectal cancer, but not in white blood cells, which may be present in stool due to acute or chronic inflammation anywhere in the gastrointestinal tract and raise ACTB levels (23)

The currently available first-generation mt-sDNA test includes information from both DNA methylation and mutational analyses. This requires the isolated DNA to be divided, with one portion of the DNA used for bisulfite treatment conversion to quantify MDMs and the other portion of unconverted DNA used for mutation assessment. Because the next-generation mt-sDNA panel does not interrogate DNA mutations, all input DNA is allocated to MDM quantification in a single reaction. It is anticipated that this modification will improve analytical sensitivity, particularly for the detection of smaller polyps which shed less DNA into the colon (28). In addition, the inclusion of just four methylation markers in the next-generation mt-sDNA panel increases the confidence in model performance by simplifying the algorithm during the development and final selection processes. Finally, DNA methylation increases with age, though the increased methylation is not consistent for every gene (29). Similar to the approach used in development for the currently approved first-generation mt-sDNA test (26), the marker selection process for the next-generation mt-sDNA panel included the assessment of age-associated methylation changes for each marker, and markers with minimal age-related variability were selected.

In addition to being highly specific, DNA methylation is more broadly informative than DNA mutations in colorectal neoplasia (30). DNA point mutations are heterogeneous and an individual mutation may only be present in a small percentage of colorectal cancer and premalignant polyps (31). Therefore, a panel with a sufficient number of mutations to detect all screen-relevant colorectal neoplasms has been estimated to require over 100 loci (32). In contrast, as few as four methylated DNA markers have been shown to have nearly 100% sensitivity and specificity in DNA extracted from primary precursor and colorectal cancer tumor specimens (25) agnostic to the molecular pathways involved in carcinogenesis (33).

There are notable limitations to this study. First, due to modifications and refinements in assay chemistry, different algorithms were used for the MSS and VSs (Fig. 1). Therefore, it is difficult to directly compare results from both studies. In addition, not all samples were collected prospectively, and sample collection was conducted 10 or more years ago; as such, samples may not be representative of the current screening population. The cross-validation of the fitted models was not as rigorous as usual. Some of the normalization constants, derived from the data, were not rederived for each random training set. A follow-up study to finalize and lock the algorithm has recently been completed using independent samples (NCT03821948, NCT03789162). This algorithm was tested in blinded analysis to establish the sensitivity and specificity of the next-generation test resulting from the novel mt-sDNA panel. Although not within the scope of this study, it is also worth noting that a lack of direct comparison between the commercially available first-generation mt-sDNA test and the next-generation mt-sDNA test is a limitation; one which we plan to address in future, more highly powered, studies. The pivotal study, “BLUE-C” (NCT04144738), was designed similarly to the clinical trial that established the performance characteristics of the first-generation mt-sDNA test (12). Subjects completed the test using the next-generation mt-sDNA panel and commercially available FIT test, followed by a screening colonoscopy. Full analysis of these data is in progress and will be reported separately.

Overall, performance of the next-generation mt-sDNA panel in the detection of early-stage colorectal cancer and prospectively collected APLs support high sensitivity and specificity for the absence of advanced neoplasia—for a next-generation mt-sDNA test. The results presented here supported additional development of the next-generation test to finalize the algorithm for validation in an independent cohort representative of the intended use population. Further evaluation of assay performance will be characterized in a prospective, multi-center clinical validation study (NCT04144738).

Z.D. Gagrat reports personal fees from Exact Sciences during the conduct of the study; in addition, Z.D. Gagrat has a patent for 9,212,392; 10,106,844; 10,954,556; 11,708,603 issued, a patent for PCT/US2020/051118 pending, and a patent for PCT/US2021/043769 pending. M. Krockenberger reports other support from Exact Sciences and Exact Sciences during the conduct of the study. B.Z. Gagrat reports personal fees from Exact Sciences during the conduct of the study; in addition, B.Z. Gagrat has a patent for Sample Collection Device pending to Exact Sciences and a patent for Sample Collector Device pending to Exact Sciences. C.M. Leduc reports other support from Exact Sciences during the conduct of the study. M.B. Matter reports personal fees from Exact Sciences Corporation during the conduct of the study. D.W. Mahoney reports other support from Exact Sciences outside the submitted work; in addition, D.W. Mahoney has a patent 11,118,228 issued, licensed, and with royalties paid from Exact Sciences. D.K. Edwards reports personal fees and other support from Exact Sciences during the conduct of the study; and is currently employed by Merck and Co., Inc. G.P. Lidgard reports other support from Exact Sciences outside the submitted work. P.J. Limburg reports other support from Exact Sciences during the conduct of the study. J.B. Kisiel reports grants and other support from Exact Sciences during the conduct of the study; other support from Exact Sciences outside the submitted work; in addition, J.B. Kisiel has a patent for 11542557 USA issued and licensed to Exact Sciences. No disclosures were reported by the other authors.

Z.D. Gagrat: Formal analysis, supervision, validation, methodology, project administration, writing–review and editing. M. Krockenberger: Formal analysis, validation. A. Bhattacharya: Formal analysis. B.Z. Gagrat: Supervision, validation, methodology, project administration, writing–review and editing. C.M. Leduc: Resources, data curation, validation, investigation, methodology, project administration, writing–review and editing. M.B. Matter: Resources, data curation, formal analysis, validation, investigation, methodology, project administration, writing–review and editing. K.D. Fourrier: Supervision, validation, methodology, project administration, writing–review and editing. D.W. Mahoney: Formal analysis, validation, writing–review and editing. D.K. Edwards V: Visualization, writing–original draft, writing–review and editing. G.P. Lidgard: Conceptualization, resources, supervision. P.J. Limburg: Resources, writing–review and editing. S.C. Johnson: Resources, supervision. M.J. Domanico: Conceptualization, supervision. J.B. Kisiel: Validation, investigation, writing–review and editing.

The authors would like to thank the following individuals: William R. Taylor (Mayo Clinic) and Hatim T Allawi (Exact Sciences Corporation) for assistance with early marker discovery; Heather Cohen, Katie Draheim, Evan Ragland, Emily Ziegler, Erica Malo, Kathy Harings, Stephanie Weber, (Exact Sciences Corporation), Dylan Braun, Austin Lynch, Carl Fergus, and Tanya Quint (formerly Exact Sciences Corporation) for sample preparation, assay execution, and data compilation and preparation; and, David A. Ahlquist (Mayo Clinic), and Marilyn Olson (Exact Sciences Corporation) for leadership and feedback.

The author Team would also like to thank Carolyn Hall, Ph.D., Feyza Sancar, Ph.D., and Will Johnson, Ph.D. (Exact Sciences Corporation)—who provided manuscript writing and editorial support. In addition, the authors would also like to thank the patients who generously participated in the original sample collection studies and the principal investigators and institutions who oversaw their enrollment.