Tumor DNA as a Cancer Biomarker through the Lens of Colorectal Neoplasia

Biomarkers can broadly be viewed as measurements providing insights in a nondestructive manner to the underlying physiologic state of a cell or an organism. The FDA defines a biomarker as “a defined characteristic that is measured as an indicator of normal biological processes, pathogenic processes, or responses to an exposure or intervention, including therapeutic interventions” (1). Biomarkers are developed for and used in both research and clinical applications and include molecular, physiologic, and radiographic characteristics. In the cancer setting, biomarkers are used for risk classification, for prevention, diagnosis and monitoring of disease, and to guide treatment selection. In this review, we focus on the detection of tumor-derived DNA as a cancer biomarker with particular emphasis on its application to early detection and management of colorectal cancer. Colorectal cancer, as the fourth most common cancer and second leading cause of cancer mortality in the United States (2, 3), is a prime candidate for biomarker application. Early detection of colorectal cancer is associated with an improved prognosis. Moreover, the long latency between the precursor polyp and invasive cancer (4–6) along with evidence from screening trials that polypectomy reduces colorectal incidence (7) demonstrates that prevention of colorectal cancer is possible. Although endoscopic screening or stool testing is effective, compliance in the general population is only 68% (8). Development of a blood-based screening test could overcome the reluctance to utilization inherent in our currently available colorectal cancer screening technologies.

Commonly, the performance of a clinical test based on biomarkers is evaluated by assessing its sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV; defined in Table 1; ref. 9). Optimal performance criteria for tumor DNA as a biomarker will vary depending on the setting in which the biomarker will be used (10). For cancer screening, where individuals not known to have cancer are screened with the goal to detect cancer early before symptoms develop, the probability that any one individual will have cancer is low. As such, a biomarker should have high specificity to avoid subjecting large numbers of healthy people to unnecessary, potentially risky, and costly medical follow-up procedures and stress. However, in a diagnostic setting, where the suspicion of cancer is high, the onus is to ensure that a cancer is not missed. In this setting, high sensitivity is more important than high specificity. Sensitivity and specificity of a biomarker must be balanced against each other, and acceptable levels will depend on the clinical context. In the screening setting, it may be preferable to sacrifice sensitivity to maximize specificity, thereby minimizing the number of patients subjected to futile diagnostic follow-up. This latter performance criterion is captured by PPV and is affected by the prevalence of disease in the tested population as well as the intrinsic performance of the biomarker test itself (Fig. 1). As others have noted, many of the concerns related to screening are due to the emphasis on maximizing sensitivity over specificity without concerted attention to the potential impact on healthy individuals (11). For colorectal cancer, reduced specificity could be less of a concern, because a prevalent strategy in the United States is to send all patients for colonoscopy. A blood test with high sensitivity but lower specificity would result in negative colonoscopic exams due to false-positive testing. However, such a screening strategy would utilize fewer resources than the alternative of colonoscopy for all.

Because genomic changes underlie cancer (12), tumor-derived DNA has many properties that make it attractive as a cancer biomarker. First, every cancer genome contains somatic alterations that distinguish it from the genome in nonneoplastic cells (12, 13). Second, somatic or acquired alterations are intrinsic and essential to “drive” the neoplastic process. However, because most somatic alterations are the result of random processes that accumulate over time, the majority of these mutations are so-called passenger mutations and are not related to the neoplastic process. Only a small subset of somatic mutations are thought to be drivers (12, 13). Third, driver mutations occur at the earliest stages of neoplasia. For example, in the case of colorectal lesions, driver mutations in APC are detected in lesions as small as a single crypt and are thought to be an initiating event (14). Somatic mutations accumulate as the tumor develops, fostering additional growth advantages and facilitating the ability to evade and disseminate (15). Fourth, as discussed in more detail below, highly sensitive and specific methods exist for the detection of cancer-associated somatic alterations. Finally, research over the past three decades has demonstrated tumor DNA in a variety of clinically accessible samples including blood (16, 17), urine (18), stool (19), sputum (20), Pap smears (21), saliva (22), cerebrospinal fluid (23), and cyst fluid (24–26).

The application of DNA as a cancer biomarker requires the ability to distinguish tumor from nontumor DNA. Fortunately, the presence of somatic DNA alterations allows such a distinction. In theory, any somatic alteration that occurs uniquely in the cancer can be used as a marker to detect tumor DNA. However, in practice, driver mutations have several advantages. First, driver mutations are more likely to be clonal (i.e., present in every tumor cell of a given tumor) than subclonal (i.e., display tumor heterogeneity; refs. 27, 28). Second, due to their neoplastic promoting properties, driver mutations are more likely than passenger mutations to be preserved as tumors advance and metastasize (27, 28). Third, driver mutations occur more commonly and are more likely to be present in tumors from different subjects than passenger mutations (12). Fourth, most driver mutations occur in a limited subset (29) of the approximately 20,000 protein coding genes in the human genome (30), greatly reducing the number of loci that need to be queried.

There are two general strategies for detection of tumor DNA in clinical samples: tumor-informed and tumor-blind. The tumor-informed strategy can be applied when tumor tissue is available and a cancer marker can be developed by utilizing the tumor tissue to sequence portions of the genome or the entire genome. Somatic mutations in passenger and driver genes can be readily identified and even private mutations can be used as markers (i.e., mutations not commonly or ever seen in other tumors). Only (a subset of) these specific somatic alterations are subsequently used to distinguish tumor DNA from DNA derived from nontumor cells. This approach is patient specific and an efficient and sensitive way to assess for the presence of minimal residual disease (MRD).

Although a tumor-informed approach has clear technical and statistical power advantages for test performance, a tumor-blind strategy must be applied when tumor tissue is not available, such as in the cancer screening setting. Most commonly, with screening, a panel of driver gene alterations characteristic of the cancers of interest are applied to accessible clinical samples, such as plasma or stool.

There are several distinct types of somatic alterations that occur in human cancer (12). Each type has different attributes with respect to sensitivity, specificity, and ease of detection and implementation. In this review, we will consider four different types of somatic alterations to distinguish tumor from nontumor DNA. In a rough order of decreasing specificity, these include: (i) rearrangements or translocations, (ii) aneuploidy or copy-number changes, (iii) single base substitutions (SBS), and (iv) small insertions and deletions (indels). Other characteristics, such as DNA length (31, 32) and fragmentation pattern (33, 34), can also be used to detect tumor-derived DNA, but these biophysical features are not considered somatic alterations in the DNA sequence. Using rearrangements or translocations for detecting tumor DNA has the potential to be maximally sensitive and specific (35–37). However, because these approaches typically must be tumor informed and bespoke for individual patients, their clinical application is limited, and they won't be discussed further. For many clinical applications, we have found that somatic SBSs, indels, and aneuploidy offer the optimal combination of sensitivity and specificity that can be pragmatically utilized for cancer screening and disease monitoring. Moreover, where specific information about driver gene alterations is required to guide clinical management (e.g., for identification of preexisting mutations that confer resistance to EGFR blockade in the treatment of metastatic colorectal cancer), SBSs/indels often provide the only solution.

Epigenetic alterations recorded as changes in the methylation status of CpG dinucleotides in DNA can also be used as biomarkers. Specific methylation changes (e.g., VIM, SEPT9, BMP3, and NDRG4) are common enough that they can often be utilized in a tumor-blind approach. Indeed, both FDA-approved DNA-based tests (Cologuard and Epi proColon) for colorectal cancer utilize methylation changes as markers. However, unlike somatic alterations, the source of detected methylation changes is less clear. They may be the result of tumor-specific methylation changes (i.e., somatic), or reflect the cell type of origin of the tumor, or be released from normal cells in the tumor bed. Thus, the developmental and age-related plasticity of epigenetic markers can be a challenge for specificity. At the same time, the ubiquitous and tissue-specific nature of epigenetic differences may enhance the detection of tumors and provide important clues to the tissue of origin (TOO) of a tumor.

DNA replication in healthy cells is not perfect, and in fact, such imperfection is essential to the evolution of life (38). Estimates for DNA replication errors suggest that many, if not most, cells will acquire private somatic mutations after a single replication cycle (13, 38). Thus, somatic alterations are not unique to cancer and occur in virtually every normal human cell. However, in the absence of the clonal expansion associated with neoplasia, these isolated nonclonal somatic mutations are not typically detectable in clinical samples because of their rarity. Still, using sensitive massively parallel sequencing (MPS) techniques, nonclonal somatic mutations have been detected in a variety of tissues and shown to increase with age (39–41). The vast majority of these mutations have no known functional consequence or occur in genes not known to promote neoplasia, but occasionally they do occur in driver genes and can be associated with various degrees of clonal expansion ranging from histologically undetectable expansions to clinically evident benign lesions and frank cancers (42–46). Although it appears that only malignant lesions typically produce enough tumor DNA to be detected at remote collection sites (i.e., blood), a nonmalignant source immediately proximate to the site of sample collection can contribute sufficient DNA for detection. These potential nonmalignant sources must be accounted for in mutation-based biomarker assays. In the case of colorectal cancer, clinical samples derived from blood and stool are associated with known nonmalignant sources of driver gene mutations. In blood samples, somatic alterations of driver genes derived from Clonal Hematopoiesis of Indeterminate Potential (CHIP) can be found in cell-free DNA (cfDNA). CHIP is a relatively common condition in the aging population due to oligoclonal expansion of mutated hematopoietic stem cells (40, 41, 47, 48). Although CHIP itself does not appear to be associated with any hematopoietic morbidity, it is associated with an increased risk of hematopoietic malignancies, cardiovascular disease, and overall mortality (41, 48, 49). As much as 10% to 25% of individuals over the age of 70 may have molecularly defined CHIP (50). Fortunately, somatic alterations attributable to CHIP are readily identified by sequencing the cellular component (i.e., white blood cells) of the blood sample and in practice are not a barrier for using somatic alterations as cancer biomarkers in blood. Nevertheless, the possibility of CHIP as a source of mutations must be considered because white blood cells are currently not routinely sequenced in many liquid biopsy applications. In stool samples, benign polyps and adenomas in particular can shed tumor DNA and share many of the same genetic alterations as those observed in colorectal cancer (51, 52). Eliminating these as sources for somatic mutation is not simple. However, because the identification of adenomas and large polyps is considered clinically beneficial and because they are relatively safely removed by colonoscopy (53), detection of somatic alterations from these lesions is favorable in the setting of colorectal cancer screening.

Detection of tumor DNA in clinical samples is technically challenging due to a variety of issues. First, most DNA in clinical samples is derived from normal cells. Surprisingly, this can even be true in tumor tissue samples, where the nonneoplastic component (e.g., inflammatory cells and stromal cells) can be greater than the neoplastic component. It is especially true for samples collected at a distance from the tumor. Second, the DNA in clinical samples such as plasma and stool is highly degraded, requiring the assays to assess relatively small fragment sizes. The average fragment size in plasma from healthy individuals is approximately 170 bp and is even smaller and more variable in patients with cancer (31, 54–58). Third, the amount of tumor DNA can be limiting in some clinical samples especially in plasma samples; a typical 1.0 mL plasma sample has approximately 4 ng of cfDNA or about 1,200 haploid genome equivalents in a healthy individual (59). Although some patients with cancer display elevated levels of total cfDNA, many do not, and the source of elevated levels is not neoplastic cells (37). Because somatic alterations are often present at frequencies less than one mutant molecule in 10,000 molecules, several milliliters of plasma are necessary for optimal assay sensitivity. To summarize, in order to employ tumor DNA as an effective biomarker, assays must be able to detect tumor DNA present at a very low fraction in limited amounts of highly fragmented DNA.

The substantive goal is to detect somatic SBSs, small indels, and aneuploidy when present in much less than approximately 1% of the analyzed cell equivalents. For SBSs and indels, this challenge corresponds to detecting mutations present at a mutant allele frequency of 0.5% or less. Early demonstrations of tumor DNA in clinical samples including stool and plasma often employed qualitative mutation-specific assays such as mutant-specific ligation (60) and mutant-specific PCR (17). These assays examined a population of both wild-type and mutant molecules and often used sensitive mutant-specific probes to determine if a mutation was present. However, because these early techniques were qualitative in nature, they often lacked the sensitivity and signal-to-noise ratio necessary to robustly detect rare somatic mutations from background mutations (Fig. 2A).

To overcome these challenges, methods relying on the discretization of a population of DNA molecules into single templates followed by scoring individual molecules, as opposed to scoring an aggregate population as wild type or mutant, were developed, which have distinct advantages in terms of quantitative precision, sensitivity, and specificity (Fig. 2A). Indeed, the first demonstration of tumor DNA in stool and urine relied on cloning individual molecules and scoring these clones by hybridization with wild-type and mutant-specific probes (18, 19). In 1999, this concept was extended and formalized with the advent of digital PCR (61). This technique obviates the need for cloning and instead determines the mutation status of single molecules by partitioning input DNA into many individual PCR reactions. Because a single template molecule is partitioned into each PCR reaction, the resulting PCR product is either 0% mutant or 100% mutant—that is, the mixture of wild-type and mutant templates can be read out as a “digital” rather than “analog” output (Fig. 2A).

This simple, yet powerful, concept underlying digital PCR ushered in a variety of new “digital” technologies that could be applied to the detection and quantification of somatic variants in clinical samples. In 2003, the throughput of digital PCR was substantially increased with the advent of BEAMing (62, 63). In this technology, template DNA is amplified within hundreds of thousands of individual emulsion reactions which convert each individual molecule into a bead particle decorated with many PCR copies which can be subsequently scored as wild type or mutant via flow cytometry. Digital droplet PCR (64) similarly discretizes the population of template DNA molecules into individual emulsion reactions, but the droplets themselves (as opposed to beads) are analogously scored.

MPS is inherently a digital technology—single molecules are partitioned into individual reactions which convert each template into a locally amplified cluster whose sequence can then be ascertained with traditional sequencing-by-synthesis chemistry (65, 66). Unlike digital PCR and its related techniques which can evaluate a single or several target mutations, MPS can query multiple bases in a high-throughput and cost-effective manner. Such an advantage is critical for cancer screening in which the identity of the particular mutation of interest is unknown and therefore the evaluation of several potential target regions across the cancer genome is required. Unfortunately, this critical advance comes at a cost—namely, the error rate of MPS is often between 0.05% (67, 68) and 1% (69, 70) which confounds its ability to accurately resolve rare mutations.

Several strategies have been developed to overcome this limitation. The use of improved base-calling algorithms and statistical techniques (71–74) can eliminate some of the errors introduced during sequencing. Similarly, the use of DNA repair enzymes can eliminate certain artifacts introduced during library preparation (75, 76). Nevertheless, even with these improvements, errors introduced by sequencing instrumentation are often limiting, particularly in the evaluation of mutations in clinical samples where the frequency of tumor DNA is often less than 1 in 10,000. A particularly powerful approach, first described in 2011 as the Safe Sequencing System (SafeSeqS; Fig. 2B), attempts to identify and correct errors, rather than eliminate them at their source (77). The SafeSeqS approach relies on use of molecular barcodes to uniquely tag and redundantly sequence DNA input molecules. The molecular tags serve to group redundant reads that originated from the same parental DNA molecule. This grouping permits distinguishing sequencing errors from true rare variants—sequencing artifacts are present in only a subset of the reads which share the same molecular barcode, whereas bona fide rare mutations are present in all or a vast majority of redundant copies. Molecular tags can either be incorporated exogenously, for example with the use of PCR primers or ligation-based adapters carrying degenerate or prespecified barcode sequences, or in the case when DNA is randomly fragmented, the genomic coordinates of the fragment breakpoints can be used as molecular tags themselves (i.e., endogenous barcodes). A variety of sequencing technologies relying on the use of molecular barcodes for error correction have been described, some using exogenously incorporated identifiers (22–25, 59, 78, 79), endogenously incorporated identifiers (39), or a combination thereof (80–83). These approaches have reduced the error rate of MPS to 10⁻⁴ to 10⁻⁵ mutations/bp and have been successfully applied to detect trace levels of tumor DNA in a variety of clinical samples (22–25, 37, 59, 78, 79, 81, 84–89).

Epigenetic differences in the form of the methylation of cytosine bases can also be used as biomarkers. Many of the same MPS approaches used to detect somatic mutations can be applied to detection of methylation differences by exploiting bisulfite conversion of DNA. In bisulfite conversion, denatured DNA is treated with sodium bisulfite which converts unmethylated cytosine into uracil while methylated cytosine (5-methylcytosine in humans) remains unchanged. Upon replication for further analysis, unmethylated cytosine bases are observed as C to T transitions relative to the expected sequence. Because of the global conversion of all unmethylated cytosines, simultaneous detection of somatic mutations is complicated for C to T transitions. The high efficiency of bisulfite conversion coupled with MPS (90) allows the sensitive detection of methylation difference at genome-wide levels (91). For evaluation of individual loci, methylation-specific primers can used to assess methylation states on bisulfite-converted DNA. In addition, alternative methods for detecting methylation differences that do not require bisulfite conversion exist. These methods typically employ methylation-sensitive restriction enzymes or specifically bind DNA containing 5-methylcytosine. Using methylation changes as a biomarker presents challenges that are distinct from those for somatic mutations. Although methylation differences are also heritable, they are more fluid and ubiquitous than genetic alterations. Accordingly, the colorectal cancer methylome has thousands of abnormally methylated genes (92), but only a subset appears to be “drivers” or functionally significant in promoting a neoplastic phenotype (93). In addition, it is important to note that the methylation changes used as markers may be either specific to cancer cells (i.e., somatic) or, more commonly, reflect the TOO. Consequently, the use of epigenetic markers to serve as indicators of TOO has been suggested as an approach for tumor localization for blood-based multicancer screening tests (discussed below).

Aneuploidy or copy-number changes provide another marker that can detect tumor DNA at low concentrations. Although a variety of methods can detect aneuploidy in tumor samples, approaches for consideration generally employ MPS to identify large numbers of fragments which are subsequently mapped to the reference human genome. The density of correctly mapped fragments should be uniform across the diploid portions of the genome, and copy-number changes are deduced by evaluating increases and decreases in this density. The first clinical applications of this approach were not in cancer but in the setting of noninvasive prenatal diagnosis (NIPD). In this setting, whole-genome libraries are produced and sequenced to low coverage by MPS. Fetal chromosome copy aberrations can be detected in the cfDNA component from the blood of pregnant women as early as the first trimester with the fetal component representing approximately 3% to 6% of the cfDNA present (94, 95). It was in the setting of NIPD that cancer-associated aneuploidy was first detected in the plasma of women with previously undiagnosed cancers (96–99). Subsequently, a variety of studies have demonstrated copy-number aberrations in the plasma of patients with cancer including patients with colorectal cancer (36, 91, 100, 101). Recently an alternative to low coverage whole-genome sequencing has been described for the detection of aneuploidy in plasma (102, 103). Rather than using a whole-genome library as a target for MPS, this approach, called Repetitive Element AneupLoidy Sequencing System (RealSeqS; ref. 103), uses a single primer pair to amplify approximately 350,000 amplicons distributed throughout the genome. Mapping of these sequences to the genome allows efficient detection of copy-number variation with similar or better sensitivity than low coverage whole-genome sequencing. As a single analyte, RealSeqS detected circulating tumor DNA (ctDNA) in the form of aneuploidy in 49% of plasma samples from eight common nonmetastatic cancers at 99% specificity (103).

Molecular barcoding eliminates virtually all sequencing errors, but artifacts introduced during library preparation (e.g., those that arise during early PCR cycles) persist (77). These low-level errors pose a challenge to confidently detect mutations present in samples at single or very low copy number, the detection of which is critical for the development of molecular diagnostics for sensitive detection of early stage disease. A particularly powerful strategy for eliminating these types of errors leverages the fact that DNA is a double-stranded molecule and inherently encodes an informationally redundant copy in its structure (80, 82, 83). As a result, bona fide mutations are present in both strands of a DNA molecule. Indeed, the cell's replication machinery capitalizes on this redundancy for high fidelity replication.

A variety of approaches have adopted this strategy to improve the accuracy of rare variant detection with MPS. For example, in Duplex Sequencing (80, 82, 83), sequencing adapters containing complementary double-stranded molecular barcodes are ligated to DNA fragment ends which permits unambiguous sequence determination of the individual strands that comprise an original DNA duplex. Error suppression is achieved by requiring mutations to be present in both strands of the original template molecules. Another approach, BiSeqS (90), utilizes bisulfite conversion to distinguish and independently sequence the DNA strands. This PCR-based strategy is particularly well-suited to assaying limited quantities of DNA (such as those found in plasma and stool) but is limited to assaying transversions and small indels. Future molecular diagnostics relying on the presence of mutant DNA as a biomarker are likely to employ variants of these sequencing strategies.

For an artifact to be introduced when sequencing both DNA strands, two criteria must be met: (i) an error must be introduced in one strand, and (ii) the same complementary error must be introduced on the other. As a result, the theoretical error rate of duplex sequencing approaches is < 10⁻¹² mutations per bp (80, 82, 83). This figure represents the technical, rather than biological, limit of detection. When applied to clinical samples, the limit of detection is significantly higher due to two factors: (i) the amount of clinically obtainable DNA and (ii) biological processes. The amount of total DNA available from clinical specimens (e.g., plasma, stool, urine, cyst fluid) is often < 33 ng (22, 23, 25, 37, 59, 78, 79). Because DNA molecules are discrete entities, the lower limit of detection is therefore bounded by the number of templates available. As a result, a sensitivity of approximately 0.01% is sufficient to detect one or two mutant molecules within the approximately 10,000 wild-type templates that are obtainable from clinical samples. Furthermore, the limit of detection is also affected by the presence of mutations arising from noncancerous biological processes. As discussed above, somatic mutations arise from normal cell replication and from nonmalignant processes (most notably CHIP in the case of blood). Like tumor DNA, these mutant molecules stemming from nonneoplastic conditions can also be detected in clinical specimens and can be a challenging confounder for detecting the presence of mutant tumor-derived DNA (104). These biological processes in combination with sample quantity, rather than technical artifacts, are likely to limit the limit of detection. In the case of blood, paired analysis of plasma and white blood cells can mitigate these challenges by readily identifying CHIP-derived variants (104, 105). It remains to be determined whether it is necessary (and if so, how) to adopt a similar strategy for other bodily fluids. Finally, limiting the analysis to driver gene mutations is expected to limit the potential confounding contribution of mutations from noncancerous biological processes. Because these genetic events are often clonal and drive tumorigenesis, focusing on these types of mutations (as opposed to passenger gene mutations) also mitigates the concern that a ctDNA-based test result could be confounded by tumor heterogeneity or evolutionary temporal dynamics.

Every biomarker has advantages and disadvantages. For example, although mutations in tumor DNA can be sensitive and specific for detecting cancer, tumor DNA in noninvasive clinical samples may be limiting or simply undetectable, especially in early stage disease. Likewise, several classic protein biomarkers (e.g., CEA, CA19–9, CA-125), which have been widely and successfully used for disease monitoring, have proven inadequate for screening because of poor specificity, sensitivity, or both. Using combinations of biomarkers or multianalyte assays is based on the premise that individually suboptimal biomarkers can be effectively used in combinations to achieve meaningful test performance. As discussed below, several of the most promising clinical applications benefit from tests employing a multianalyte approach.

The possibility of using DNA released from tumor cells as a biomarker in stool was first demonstrated nearly 30 years ago, with the identification of KRAS mutations in DNA isolated from the stool of patients with colorectal cancer (19). From a detection perspective, tumor DNA identified in the stool must be distinguished from both an excess of DNA derived from nontumor cells and the even vaster excess of DNA from bacteria that inhabit the colon. The multitarget stool DNA (mt-sDNA) test (Cologuard, EXACT Sciences) combines a fecal immunochemical test (FIT) for human globin with stool testing for KRAS mutations and for aberrant methylation in the promoter regions of BMP3 and NDRG4 (106). In a study comparing one-time use of mt-sDNA to FIT in nearly 10,000 average risk adults who underwent colonoscopy as the gold standard, the sensitivity of mt-sDNA for colorectal cancer was superior to FIT (92.3% vs. 73.8%, P = 0.002), although mt-sDNA was less specific (86.6% vs. 94.9%, P < 0.001; ref. 106). Over 3 million Cologuard tests have been performed worldwide (107).

There is currently one FDA-approved blood test for early detection of colorectal cancer (Epi proColon, Epigenomics AG) which is based on the detection of methylated SEPT9 in plasma. A meta-analysis of 14 studies with SEPT9 showed a pooled sensitivity of 67% and specificity of 89% for colorectal cancer, albeit with significant heterogeneity (108), and a meta-analysis that included 25 studies performed similarly (109). However, a test with a sensitivity of around 70% with over 10% false positive findings may be of limited utility in some clinical applications, given the large numbers of false positive in proportion to true positive tests (Fig. 1).

The Early Detection Research Network (EDRN; https://edrn.nci.nih.gov/) has assisted in pioneering a plasma-based combined ctDNA and protein panel multicancer test named CancerSEEK (Fig. 3A). CancerSEEK queries 61 mutations/amplicons in 16 genes (59) and utilizes SafeSeqS. Because not every cancer releases detectable levels of ctDNA, eight selected proteins are also assessed as part of the test to further enhance performance. A case–control study of 1,005 patients with cancer with eight types of stage I to III cancer including ovary, liver, stomach, pancreas, esophagus, colorectal, lung, and breast and 812 controls was performed. A single composite CancerSEEK score for each individual was derived by combing a log ratios score for mutations with the eight protein biomarkers using a logistic regression algorithm. Using a 10-fold cross-validation study design, the median sensitivity across the eight cancer types evaluated was 70% at >99% specificity. The sensitivity ranged from 98% in ovarian and hepatocellular cancer to 33% in breast cancer; for colorectal cancer, the sensitivity of CancerSEEK was 66% (59). Because CancerSEEK was designed as a multicancer test and the mutations and individual protein markers are not specific for a particular tumor type, the possibility of identifying tumor type from the liquid biopsy results was explored. A machine-learning algorithm using ctDNA mutation metrics, 39 protein biomarker levels, and patient sex as input was able to localize the source of the cancer to one of two anatomic sites in a median of 83% of patients (59).

Consistent with the value and extensible nature of multianalyte approaches, addition of aneuploidy as measured by RealSeqS to the above CancerSEEK analytes increased sensitivity for seven of the eight cancers tested while maintaining 99% specificity based on 812 healthy controls (103). Likewise, sensitivity for stage I to III colorectal cancer increased from 66% to 77% at the same specificity (refs. 59, 103; Fig. 3B).

In a similar manner, machine learning was used to combine aneuploidy and DNA fragmentation patterns to create a blood test named DNA evaluation of fragments for early interception (DELFI), for multicancer detection and localization (110). Further coupling DELFI with somatic mutations detected 90% of 109 stage I to III cancers at a specificity of 98% based on a small number (n = 245) of healthy controls (110).

In another variation, a multicancer blood test was reported that employed machine learning to analyze methylation patterns in cell-free plasma DNA for both cancer detection and TOO identification (111). The machine learning algorithm utilized was trained in data from 3,052 subjects (1,531 cancer and 1,521 noncancer) and validated on 1,264 subjects (654 cancer and 610 noncancer). Stage I to III sensitivity was approximately 67% for 12 cancer types (anus, bladder, colon/rectum, esophagus, head and neck, liver/bile-duct, lung, lymphoma, ovary, pancreas, plasma cell neoplasm, stomach) and approximately 44% for all cancer types. TOO was predicted in 96% of cases with cancer-like signal in the validation set and was 93% accurate among these cases resulting in correct identification of TOO in approximately 89% of cases. In the context of this review, it is important to note that the source of the methylation signal for both detection and localization may come from tumor DNA (i.e., DNA from neoplastic cells) and/or nontumor DNA in the TOO (i.e., DNA from surrounding nonneoplastic or stromal cells).

With the growing number of studies reporting the potential of multicancer blood tests (59, 89, 103, 110, 111), there is a critical need to begin to address the safety and efficacy of such tests in healthy individuals. A recent study (DETECT-A) evaluated the safety and feasibility of multicancer blood testing coupled with PET-CT imaging to detect cancer in a prospective, interventional study of 10,006 women not previously known to have cancer (112). The multicancer blood test utilized in this study was an early version of CancerSEEK described above but did not employ machine learning to maximize sensitivity and specificity. The combination of blood testing and imaging achieved a sensitivity of 27.1% at a specificity of 99.6%. From a clinical perspective, it doubled the number of cancers detected in this population over standard of care screening and only 0.22% of the screened individuals underwent a futile invasive diagnostic procedure.

In addition to blood-based tests to detect multiple cancers, ctDNA tests for focused detection of colorectal cancer using machine learning are also in development (113). The application of blood-based testing to colorectal cancer screening holds promise for increasing compliance given the burden of colonoscopy or stool-based examinations. However, endoscopic testing, by removing precursor adenomatous polyps, has a potent effect on preventing cancer. Estimates suggest that reducing cancer incidence accounts for two thirds of the mortality reduction produced by screening (114). It is unknown whether ctDNA testing will detect precursor lesions such as advanced adenomas. Stool-based testing such as FIT or mt-sDNA only detects 25% to 42% of such lesions (106). Thus, although blood-based screening may succeed in increasing the population prevalence of testing, the preventive benefits of endoscopy may be compromised if such blood tests replace rather than supplement standard-of-care screening.

Uses of ctDNA detected in liquid biopsies as a prognostic factor and for the detection of MRD are two closely related clinical applications (detection of MRD is a form of prognostication). Several studies support ctDNA as a prognostic marker for subsequent colorectal cancer outcome. In metastatic disease, a higher ctDNA level is a poor prognostic sign and is associated with shorter survival (37). In stage IV colorectal cancer, ctDNA is a prognostic marker of therapeutic response (87). In a study of 53 patients, a ≥10-fold reduction in ctDNA pre cycle 2 chemotherapy predicted a progression-free survival of 14.7 versus 8.1 months (87).

ctDNA can be used as a measure of MRD and as a prognosticator after treatment. In a study of postoperative patients with stage II colorectal cancer who did not receive chemotherapy, 11 of 14 (78.6%) patients who were ctDNA-positive recurred compared with 16 of 164 (9.8%) who were ctDNA-negative, for a HR of 18 [95% confidence interval (CI), 7.9–40] for recurrence with a positive ctDNA test (78). In a study of 58 subjects in Sweden with stage I to III colorectal cancer, 10 of 13 patients with a postoperative positive ctDNA test (77%) recurred versus zero out of 45 (0%) with a negative test (88).

The prognostic utility of ctDNA in stage III colorectal cancer was evaluated in 96 subjects in Australia (85). ctDNA was observed postoperatively in 20 of 96 subjects (21%). The recurrence-free interval at 3 years was 47% among those positive for ctDNA versus 76% among those negative for ctDNA; the HR for recurrence was 3.8 (95% CI, 2.4–21, P < 0.001; ref. 85). After chemotherapy, 15 of 88 (17%) were ctDNA-positive. At 3 years, there was a 30% recurrence-free interval in those who were ctDNA-positive versus 77% in those who were ctDNA-negative for an HR of 6.8 (1.1–15.7, P < 0.001; ref. 85). Subjects positive for ctDNA postoperatively who convert to negative compared with those who remain positive had a more prolonged recurrence-free interval. Similarly, those negative for ctDNA who convert to positive had a shorter recurrence-free interval than those who remain negative (85).

A commercially developed “Signatera” ctDNA assay (Natera) was tested in patients with stage I to III colorectal cancer in Denmark (115). The assay targets 16 clonal somatic mutations. The sensitivity of a similar type assay was able to detect a variant allele frequency as low as 0.01%, or one mutant haploid genome in a background of 10,000 normal haploid genomes (116). At postoperative day 30, 10 of 94 patients were ctDNA-positive. Seven out of these 10 (70%) recurred, whereas 10 out of 84 (12%) of the ctDNA-negative patients recurred, with an HR of 7.2 (95% CI, 2.7–19, P < 0.001). Subjects who persistently maintained ctDNA-negative status postoperatively and during treatment had a markedly improved recurrence-free survival compared with those with ctDNA present (HR ctDNA-positive vs. ctDNA-negative: 43.5, 95% CI, 9.8–193.5, P < 0.001; ref. 115).

ctDNA is a better predictor of recurrence than CEA in stage II colorectal cancer and had a longer lead time to radiologic recurrence: 167 days with ctDNA versus 61 days with CEA (P = 0.04; ref. 78). In stage III colorectal cancer, CEA elevation was associated with a shorter recurrence-free interval, but when included in a combined analysis, CEA did not add additional predictive accuracy beyond ctDNA (85). In locally advanced rectal cancer, ctDNA added prognostic value to patients with a nonelevated postoperative CEA (HR 8.8, P < 0.001; ref. 86).

The above studies demonstrate the prognostic potential of ctDNA but have not established whether these biomarker tests effectively inform clinical care. Because of the uncertainty of clinical benefit, there is no consensus as to which, if any, patients with stage II disease should receive adjuvant chemotherapy. In stage II colorectal cancer, ctDNA status was more discriminatory in predicting the risk of recurrence than a classification based on clinical risk factors of low risk (N = 129) or high risk (N = 49) for recurrence (78). Hence, ctDNA may be better at appropriately triaging patients with stage II colorectal cancer to the need for receipt of adjuvant chemotherapy. In an ongoing Australian DYNAMIC study (ACTRN12615000381583), patients with stage II colorectal cancer are randomized to a standard management group where clinical characteristics are used to decide on whether adjuvant treatment will be employed versus a ctDNA management group. In the ctDNA management group, those who are ctDNA-positive postoperatively are treated with chemotherapy and those who are ctDNA-negative are observed. The primary objective of the study is to demonstrate that a ctDNA-based strategy will reduce the number of patients receiving adjuvant chemotherapy without compromising recurrence-free survival. A similar type study in stage II colorectal cancer is underway in the United States (COBRA, NCT-04068103), and multiple trials in Europe are planned or in process (e.g., IMPROVE-IT EudraCT#2018–00070–30, MEDOCC-CrEATE, Netherlands Trial Register# NL6281/NTR6455). A DYNAMIC study in stage III colorectal cancer in Australia is using ctDNA to tailor chemotherapy management (ACTRN12617001566325).

In a cohort of 159 patients with locally advanced rectal cancer, detection of ctDNA after chemoradiation (HR 6.6, P < 0.001) or after surgery (HR 13.0, P < 0.001) was associated with an inferior recurrence-free interval (86). These data suggest that ctDNA status may be useful to triage a decision regarding the benefit of adjuvant therapy in rectal cancer. A DYNAMIC study randomizing patients with rectal cancer to a ctDNA informed group that will determine adjuvant therapy is underway (ACTRN12617001560381).

ctDNA characteristics can also be used to personalize treatment, as study of ctDNA mutations can identify the causes that explain drug resistance. For example, ctDNA can identify the mechanism for resistance to EGFR blockade (37). In persons with non–small cell lung cancer, ctDNA is used to determine whether an EGFR T790M resistance mutation has developed. Identifying this resistance mutation can help physicians choose appropriate treatment (117).

The development of biomarkers is challenging. Issues with current biomarker development center on the emphasis on sensitivity over specificity (11) and the reliance on single markers to achieve clinical performance. The increasing use of combinations of analytes and the enhanced focus on specificity will hopefully lead to tests with better clinical performance and broader applicability. Already, detection of tumor DNA in clinical samples as a single analyte or in combination with others has shown promise for early detection, prognostication, detection of MRD, monitoring, and treatment selection. These applications will improve as the underlying technology for detection of tumor DNA improves. Of these applications, early detection has the greatest potential to reduce cancer morbidity and mortality.

With the growing number of studies demonstrating the potential of multicancer detection using blood samples from patients with previously diagnosed cancer (59, 103, 110, 111), there is a pressing need to evaluate these tests in prospective studies in individuals without a current diagnosis of cancer. Prospective studies with a large number of cases and an even larger number of controls are necessary to determine the true performance and clinical utility of these tests as sensitivity is likely to be lower for asymptomatic cancers and specificity is likely to be lower in subjects undergoing screening due to concurrent comorbidities. Moreover, an interventional study that includes long-term follow-up is the only way to judge risk versus benefit.

Finally, further studies are necessary to determine the optimal methods for tumor localization when using multicancer blood tests. Although molecular TOO determination is promising, in the absence of a perfect detection and TOO testing, patients may embark on a prolonged diagnostic journey. For example, even with a well-performing test (99% specificity), most subjects who test positive will be falsely positive (Fig. 1). Therefore, additional testing may be necessary to distinguish between a false positive test and a true positive with incorrect identification of the TOO. Likewise, even in the case of a true positive, the possibility of an incorrect TOO could result in pursuit of multiple follow-up tests or ignoring an existing cancer. These considerations raise the question of whether whole body imaging might provide a more facile tumor localization. Although historically such imaging approaches have often resulted in high rates of incidental findings, their use following a positive multicancer blood test might result in a favorable performance profile.

A growing number of studies demonstrate that tumor DNA has clinical utility as a biomarker across the spectrum of cancer clinical practice. These studies have been fueled by advances in technology that increased the ease, sensitivity, and specificity of tumor DNA detection in clinical samples. Although there are currently only a few FDA-approved tests that include released tumor DNA as a biomarker (e.g., Cologuard, Epi proColon, cobas EGFR Mutation), this number is likely to grow over the coming years due to technical advances and the growing number of promising translational and clinical studies. Ultimately, the true measure of tumor DNA as a biomarker will be the extent to which it improves early detection, and the management and outcomes of patients with cancer.

J.D. Cohen reports personal fees from Guidepoint and Gerson Lehrman Group outside the submitted work, as well as a patent for CancerSEEK licensed and with royalties paid from Thrive Earlier Detection. B. Diergaarde reports grants from NIH–NCI during the conduct of the study. N. Papadopoulos reports grants from NCI during the conduct of the study; personal fees and other from Thrive Earlier Detection (equity; member of the board of directors) and other from PGDx (equity), Cage (equity), and NeoPhore (equity) outside the submitted work; and a patent for SafeSeqS pending, issued, licensed, and with royalties paid from multiple companies and a patent for CancerSEEK pending, licensed, and with royalties paid from Thrive Earlier Detection. K.W. Kinzler reports grants from NCI during the conduct of the study; personal fees and other from Thrive Earlier Detection (equity), personal fees from Sysmex and Eisai, and other from PGDx (equity), Cage (equity), and NeoPhore (equity) outside the submitted work; and a patent for SafeSeqS pending, issued, licensed, and with royalties paid from multiple companies and a patent for CancerSEEK pending, licensed, and with royalties paid from Thrive Earlier Detection. No potential conflicts of interest were disclosed by the other author.

The authors thank Christopher Douville for providing additional details on the performance of RealSeqS on colorectal cancer. Our work in this area has been supported by The Marcus Foundation; Lustgarten Foundation for Pancreatic Cancer Research; The Virginia and D.K. Ludwig Fund for Cancer Research; The Sol Goldman Center for Pancreatic Cancer Research; The Conrad R. Hilton Foundation; and NIH grants U01-CA152753 (to R.E. Schoen), U01-CA230691 (to N. Papadopoulos), and P50-CA062924 (to K.W. Kinzler).