Circulating tumor DNA (ctDNA) as a pan-cancer screening test: is it finally on the horizon?

Assays based on mutation detection

Using a highly sensitive sequencing method known as targeted error correction sequencing (TEC-Seq), Phallen et al. [10] compared the ctDNA mutational profile of 58 cancer-associated genes from 200 patients with an established diagnosis of stage I/II cancer and 44 healthy controls. Of 200 patients with cancer, 134 (67%) had mutations detected in their ctDNA. However, for patients with stage I disease, mutations were detected in only 35/64 (54%) of cases. Sensitivity was highest in patients with colorectal cancer (83%), followed by ovarian cancer (71%), lung cancer (62%) and breast cancer (56%). For patients with stages I or II disease, sensitivity was 71% in patients with CRC, 68% in ovarian cancer and 59% in both lung and breast cancer. Importantly, there were no mutations detected in driver genes associated with solid cancer (genes responsible for driving tumor cell growth) in any of 44 healthy controls investigated. However, 16% of healthy controls had mutations in genes associated with a benign condition known as clonal hematopoiesis of indeterminate potential (CHIP) (see below).

In an attempt to further enhance sensitivity, Cohen et al. [11] developed a combined assay that detects both DNA mutations and protein biomarkers. This test, known as CancerSEEK, detects mutations at 2,001 locations across 16 cancer-associated genes as well as the concentration of eight protein biomarkers. In a retrospective case-control study that included 1,005 patients with clinically detected cancers of the ovary, liver, stomach, pancreas, esophagus, colorectum, lung, or breast. CancerSEEK was found to be positive in a median of 70% of cases. Similar to the situation with standard protein biomarkers, sensitivity increased with increasing stage of the cancer, i.e., median sensitivity was 43, 73 and 78% for patients with stages I, II and III disease, respectively. Sensitivity was highest in patients with ovarian or liver cancer (98%) and lowest in breast cancer (33%). For patients with stage I disease, sensitivity was 100% in those with liver cancer but was only 20% in those with esophageal cancer. Specificity was >99%, i.e., only 7/812 apparently healthy controls gave a positive result. Some of these subjects may subsequently have developed cancer but this does not appear to have been investigated. In the 626 patients who tested positive, use of a machine learning algorithm that include the CancerSEEK analytes and patient sex narrowed the tumor location to either of two organs in 83% of patients and correctly located the specific site in 68% of cases. In an update on CancerSEEK, Douville et al. [12] showed that a combination of aneuploidy (copy number changes) with mutation detection and the eight protein biomarkers yielded a median sensitivity of 80% for cancer detection, while only 1% of 812 healthy subjects tested positive.

An earlier version of the CancerSEEK test in combination with PET-CT imaging, was recently investigated in a prospective exploratory study known as DETECT-A [13]. Of 9,911 asymptomatic women screened, 490 (4.9%) gave a positive result in the initial baseline blood test. Following exclusion of CHIP-derived mutations, 134 subjects (1.35% of the 9,911 participants) had their initial result confirmed in a follow-up test. Of these 134 women, 127 underwent PET-CT scans or other imaging procedures. Resulting from these tests, 26 cancers were detected of which 17 (65%) were described as having early disease. In addition to these 26 cases detected using ctDNA, 24 further cancers were detected by standard screening tests and 46 by other diagnostic modalities during the study. Overall, the ctDNA test had a sensitivity of 27.1%, a specificity of 98.9%, a positive predictive value (PPV) of 19.4% and a negative predictive value (NPV) of 99.3%. Combining the ctDNA test with any form of imaging increased specificity to 99.6% and the PPV to 40.6%, but had little impact on the sensitivity or NPV. It should be stated that this study was not designed to determine whether detecting and treating early cancer reduced mortality but to show that ctDNA testing was feasible and could be safely used to detect early cancer in asymptomatic subjects.

In 2019, CancerSEEK received Breakthrough Device designation from the U.S. Food and Drug Administration for the detection of genetic mutations and proteins associated with pancreatic and ovarian cancers (https://hub.jhu.edu/2019/06/03/cancerseek-blood-test/). Further validation of CancerSEEK is ongoing in a prospective, observational study (ASCEND) involving 1,000 subjects with known or suspected cancer and 2,000 subjects with no evidence of cancer (ClinicalTrials.gov Identifier: NCT04213326). The primary objective of this study is to develop and validate the CancerSEEK test for the early detection of cancer.

Assays based on altered methylation patterns

Compared to normal healthy tissue, DNA in tumors exhibits aberrant methylation. In particular, there is a progressive gain of promoter CpG-island hypermethylation and a loss of CpG methylation in non-CpG-island promoters [14], [15]. In at least some situations, these alterations occur early in the formation of malignancy and appear to be tumor-type specific [14], [15]. As with mutant DNA, aberrantly methylated DNA can be detected in blood and thus potentially used for the early detection of cancer [16]. Advantages of measuring methylation alterations over mutations include greater abundance and enhanced tissue-specificity [17], [18], [19]. The enhanced tissue-specificity should make methylation-based assays superior to mutation detection for the identification of the tissue of origin of newly-detected cancers.

One of the most high profile although not necessary the most accurate plasma methylation tests for early cancer detection is the GRAIL test (also known as the Galleri test). This test uses targeted whole genomic bisulfite sequencing of plasma DNA and machine learning to analyze a panel of >100,000 informative methylation loci. GRAIL has been investigated in a prospective multicentre case-control study containing 6,689 individuals (2,482 patients with cancer and 4,207 apparently healthy subjects) [20]. The 6,689 patients were divided into a training (n=4,720) and an independent validation set (n=1,969). In the validation set, the overall sensitivity of the test in 12 pre-specified cancer types for patients with stages I to III disease was 67.3%. Positivity increased with increasing cancer stage, i.e., 39% in stage I, 69% in stage II, 83% in stage III, and 92% in stage IV. Specificity was 99.3%, i.e., a 0.7% false-positive rate. In contrast to the 12 pre-specified cancer types, sensitivity across a broader panel of more than 50 different cancer types was 18% in stage I, 43% in stage II, 81% in stage III, and 93% in stage IV. In patients with a diagnosed cancer, the tissue of origin was correctly localized in 93% of cases.

This test is undergoing further validation in several different clinical trials. One of these trials (known as STRIVE) is a prospective multi-center observational cohort study aiming to validate GRAIL for the early detection of breast and other cancers (ClinicalTrials.gov Identifier: NCT03085888). Another trial (SUMMIT) involves its validation for detecting multiple cancers at an early stage (NCT03934866). The third trial (The PATHFINDER) is a prospective, multi-center study assessing the possible implementation of the test into clinical practice (ClinicalTrials.gov Identifier: NCT04241796). In addition to these trials, it is planned to pilot the Galleri test in 165,000 subjects in England (140,000 people aged 50–79 years who have no signs or symptoms of cancer and 25,000 people aged 40 years and above with symptoms of cancer (https://www.england.nhs.uk/2020/11/nhs-to-pilot-potentially-revolutionary-blood-test/).

The GRAIL test was recently granted an US FDA breakthrough device designation (http://www.pmlive.com/pharma_news/grails_multi-cancer_blood_test_gets_fda_breakthrough_status_1287854). It is important to state however that the Galleri test is intended to be complementary to, and not a replacement for U.S. guideline-recommended cancer screening tests.

A different plasma DNA methylation test, known as PanSeer, targets 10,613 CpG methylation sites across 477 genomic regions [21]. In a study involving 113 patients with an established diagnosis of cancer and 207 healthy controls, PanSeer was positive in 88% of patients at 96% specificity. Sensitivity was claimed to be similar in patients with early and late stage cancers. In 191 asymptomatic subjects who later developed one of five different cancers (stomach, esophageal, liver, CRC, lung), the test had a sensitivity of 95% at 95% specificity. Positivity was found up to 4 years in advance of diagnosis.

A further ctDNA methylation test, known as cfMeDIP (cell-free methylated DNA immunoprecipitation and high-throughput), uses affinity capture methods to enrich plasma DNA for densely methylated CpG regions of the genome. Following sequencing of the captured fragments of DNA, cfMeDIP employs machine-learning algorithms to differentiate patients with cancer from healthy controls [22]. To date, cfMeDIP has been mostly investigated in patients with renal and brain cancer. In a study involving 99 patients with renal cancer and 28 healthy controls, the test gave an AUROC (area under the receiver operating characteristic) curve of 0.99 [23]. For differentiating between patients with renal and bladder cancers (n=21), the AUROC was 0.97. In a further study, cfMeDIP was found to differentiate between patients with gliomas (n=59) and those with other cancer types or healthy controls (n=388) with an AUROC of 0.99 [24]. Furthermore, the test was able to distinguish between different types of brain tumors that can look similar on imaging such as gliomas, meningiomas, glial-neuronal tumors and brain metastases from systemic cancer.

Assays based on different ctDNA fragment size

While plasma DNA fragments derived from normal cells shows a predominant (modal) peak size of 167 bp, ctDNA tends to exist in both shorter (90–150 bp) and longer fragment sizes (250–300 bp) [25], [26]. Mouliere et al. [25] were one of the first to exploit this difference in fragment size to differentiate between patients with cancer and healthy controls. By selecting fragments between 90 and 150 bp, these authors were able to identify clinically important mutations and copy number alterations that would have been missed in the absence of sized-based separation. Integrating fragment length and copy number analysis of plasma DNA, led to a greater number of metastatic cancer detected, i.e., area under the curve (AUC) was >0.99, compared to AUC < 0.80 without incorporating the fragmentation patterns.

Subsequently, Cristiano et al. [27] exploited this fragment size difference in an attempt to develop an early diagnostic test for cancer. The test known as DELFI (DNA evaluation of fragments for early interception) uses whole genomic sequencing and artificial intelligence to detect different patterns of DNA fragmentation in plasma from healthy controls and cancer patients. Using low coverage sequencing, the diagnostic potential of DELFI was investigated in 208 patients with seven different cancer types and in 215 apparently healthy individuals. Following the application of a machine-learning model that combined DNA fragmentation patterns and copy number changes, the test was found to differentiate between healthy subjects and patients with different cancer types with an overall sensitivity of 73% at 98% specificity. Sensitivity varied from 68% in stage 1 disease to 79% in those with stage III disease. Lowest sensitivity was found in breast cancer (57%) and highest in lung cancer (100%). The addition of ctDNA mutations to the DELFI data increased the sensitivity for cancer detection from 73 to 91%, while maintaining similar specificity. The test however, was poor in correctly identifying the tissue-of-origin of the cancers, i.e., it was able to identify the tissue of origin to one of two sites in only 75% of cases.

Specificity of ctDNA tests

Compared with existing protein biomarkers, tumor-derived DNA might be expected to be more cancer-specific. Indeed, the emerging results with Cancer SEEK and GRAIL mentioned above, reported specificities of >99% using samples from apparently healthy controls. This high specificity may initially appear impressive and although no direct comparisons were made, it is likely to be superior to that available from standard protein biomarkers.

Although mutant DNA might be expected to be rarely detected in the circulation of healthy subjects, certain benign lesions, especially CHIP, can release mutant DNA into the circulation [28], [29]. CHIP involves the clonal expansion of white blood cells and occurs in approximately 10–20% of apparently healthy subjects over 70 years of age [30]. Using deep sequencing, Razavi et al. [31] reported that most of the DNA mutations in blood whether form healthy control subjects or patients with cancer had characteristics of been derived from CHIP (81.6% in controls subjects and 52% in patients with malignancy). In another study, CHIP-derived mutations were detected in 76% of patients with lung cancer and in 91% of high-risk controls [32]. In this study the VAFs (the relative frequency of a variant/mutation expressed as a percentage) of the mutations detected in plasma and leukocytes were significantly correlated, suggesting a blood cell origin for the detected mutations. CHIP mutations with the highest VAF occurred in DNMT3A, TP53, TET2, ASXL1, PPM1D and IDH1/2 [28].

To overcome the problem of CHIP-derived mutations leading to false-positive findings, several strategies have been proposed. These include parallel measurement of white cell DNA or performing sequencing on size-based selection of circulating DNA fragments [32], [33]. As mentioned above, in the DETECT-A study [13], mutations related to CH were excluded by simultaneously sequencing DNA from the white blood cells. Performing these additional procedures however, is costly and time-consuming, thus limiting the attraction of ctDNA as a population-based screening test for cancer.

As well CHIP, mutations have been detected in several other benign lesions and even in normal tissues [34], [35]. Thus, following a comprehensive analysis of approximately 6,700 normal samples across 29 normal tissues, Iizhak et al. [34] identified multiple somatic mutations in a range of different apparently normal tissues, especially in sun-exposed skin, esophagus, and lung. These mutations tended to accumulate over both time and number of cell divisions. Clearly, therefore, DNA mutations are not unique to malignant tissue [34], [35] and indeed may be the “new normal” [34].

These findings beg the question whether benign lesions other than CH or even healthy individuals release mutant DNA into the circulation. The answer to this question appears to be yes. For example, in a study involving subjects from Sub-Saharan Africa where that the hepatocarcinogen, aflatoxin contaminates much of the traditional diet, mutation in TP53 were detected in 15% of patients with liver cirrhosis and in 3% of apparently healthy individuals [36]. In another study, mutations in KRAS or TP53 were found in plasma from apparently healthy subjects in 13/1,098 (1.2%) and in 20/550 (3.6%), respectively. Of these mutation-positive subjects, 16 developed cancer, on average after 18.3 months of follow-up [37].

Sensitivity of ctDNA tests

As well as high specificity, high sensitivity for early stage tumors is an important criterion for a cancer screening test. As with standard protein biomarkers, the sensitivity of ctDNA varies with cancer stage and site of origin. Thus, in the studies reported to date, sensitivity increased with increasing stage, i.e., from approximately 40% in stage I to approximately 80% in stage III malignancy [12], [20]. The most likely reason for the greater sensitivity with higher stage disease is that increasing amounts of mutant DNA are shed into the circulation as the tumor size/volume enlarges. In addition to stage, sensitivity depends on the cancer type. Thus, using plasma from 21,807 patients with advanced tumor of diverse origins, the prevalence of ctDNA ranged from 51% in glioblastoma to 93% in small cell lung cancer [38]. As regards, levels of ctDNA, these were highest in CRC and lowest in gliomas [38].

Similar to the situation with specificity, direct comparisons of the comparative sensitivity of ctDNA and standard biomarkers have not been performed. Nevertheless, the above preliminary findings suggest that ctDNA has enhanced sensitivity compared to existing protein biomarkers, at least for some tumors. For example, the CancerSeek test was positive in 98% of patients with ovarian cancer at a specificity of >99% [11]. This accuracy of CancerSeek appears to be higher than that of the standard biomarker for ovarian cancer, i.e., CA 125 [39]. For example, following a systematic review of the literature followed by a meta-analysis, Ferraro et al. [38] calculated that CA 125 had an overall sensitivity of 79% and specificity of 78% for ovarian cancer. The GRAIL test in patients with pancreatic cancer which was reported to be positive in 63% of patients with stage I disease, 83% in stage II, 75% in stage III and 100% in stage IV [20], also appears to be more sensitive than the standard biomarker, for pancreatic cancer, i.e., CA 19-9 [40]. However, neither CA 125 for ovarian cancer or CA 19-9 for pancreatic when used alone can be recommended as screening tests for their respective cancers because of their low prevalence and thus low PPV in asymptomatic subject (see below).

While the emerging but indirect data suggests that ctDNA has superior sensitivity than established biomarkers for detecting cancer, it ability to detect stage I disease (<2 cm) is limited, i.e., a sensitivity of 43% with CancerSEEK [11] and 39% with GRAIL [20]. Potential strategies to improve sensitivity include drawing larger volumes of blood (not always feasible), using ultra-low sensitive sequencing, increasing the number of mutations analyzed, combining different types of genetic alterations (mutations, copy number changes and methylation patterns), combining ctDNA measurements with protein biomarkers (as in CancerSEEK) or enriching for low molecular weight DNA fragments prior to mutational analysis [25], [26], [27]. Again, these additional requirements limit the attractiveness of ctDNA assays for cancer screening.

Although potential strategies are available for enhancing sensitivity, calculations by Fiala and Diamandis [41], [42] indicated that ctDNA has a low probability of detecting tumors with a diameter of ≤10 mm using 10 ml of whole blood. For tumors with a diameter of ≤6 mm, these authors concluded that it was not possible to detect ctDNA in 10 ml of blood, irrespective of the sensitivity or reliability of the assay used [41], [42]. At this small size, some tumors may already have metastasized and thus be potentially incurable by surgery but they might be curable using surgery followed by adjuvant therapy, especially if the amount of metastasis is minimal.