The principles of screening

In 1968 Wilson and Jungner published a set of principles to guide the selection of diseases amenable to screening2 that are largely considered the gold standard in the assessment of screening decisions. These principles have been reevaluated more than once (online supplemental table S2), as technology has made it easier to detect diseases, including cancers, in earlier, preclinical stages using molecular diagnostic tests that characterize genomic rather than pathological alterations. Here we apply the most recent iteration of the principles of screening31 as early detection tests come of age. We will focus primarily on the disease and test principles (the principles for screening programs are beyond the scope of this article; see online supplemental table S3).

The principles state that the disease or condition, in this case cancer, should be an important health problem, with well-understood epidemiology and natural history, a detectable preclinical phase, and clearly defined target population for testing. A prudent approach with any new technology would be to use it for the group most likely to benefit—in this case, those who are most likely to develop cancer. Doing so intrinsically improves the potential ratio of benefits to harms. Once evaluations in targeted elevated risk populations prove successful, incremental expansions of the target population can be undertaken informed by risk and harm data. This is consistent with guidance from the WHO regarding the use of pilot screening programs to gather real-world data to inform more widespread screening based on evidence.32 ³³ The United States Preventive Services Task Force already recommends screening for four cancer types (lung, colorectal, breast and cervical), so applying new screening technologies to one or more of these cancer types is a reasonable initial approach.

According to the principles, a cancer screening test should be fit for purpose, have good performance characteristics, be acceptable and affordable to the target screening population, and have clinical utility. The emergence of new technologies for the analysis of cfDNA circulating in the blood holds great promise to improve performance and access of current screening and may enable screening for other cancers for which no approaches are currently available. Because the amount of tumor-derived ctDNA present in patients with early-stage cancer can be relatively small compared with the total amount of cfDNA, looking for evidence of cancer in the blood of asymptomatic individuals is very much a needle-in-a-haystack problem. Methods and techniques have become available to find that ‘needle’; the challenge is to do so with robust performance characteristics in early-stage cancers and at affordable costs.

Today, there are three basic approaches to the detection of cancer by analyses of cfDNA (table 1). These approaches include testing for (1) somatic mutations, (2) methylation profiles, or (3) whole-genome fragmentation profiles (figure 1).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Liquid biopsy approaches for cancer detection. Cell-free DNA technologies evaluating mutations or methylation are often costly, multistep, time consuming processes prior to data analyses. Low-coverage whole-genome sequencing approaches can minimize arduous laboratory protocols at low-cost with broadly applicable bioinformatic pipelines. cfDNA, cell-free DNA; MAF, mutant allele frequency.

Cancer screening using somatic mutations

DNA sequencing of tumor DNA has been widely used in the development of companion diagnostic tests to select appropriate targeted therapies for individual patients. The great majority of Food and Drug Administration-approved companion diagnostic tests are based on sequencing tumor samples for the presence of specific mutations (eg, EGFR T790M mutation) to guide treatment selection.33 Use of mutation-based ctDNA detection approaches for screening asymptomatic individuals, however, has not gained traction, largely because the amount of ctDNA is very small (often <1% of the total cfDNA) in early-stage cancer. Such an approach to asymptomatic screening would necessitate sequencing very deeply (>10,000×) and using error-reducing sequencing approaches (eg, molecular identifiers combined with redundant sequencing and consensus calling) to reliably determine if rare mutated-DNA fragments are present in cfDNA. Using whole-genome sequencing for such approaches would be prohibitively expensive. To get a sense of the challenge, DNA sequencing for a specific mutation at 60,000-fold unique coverage depth will only reveal on average 30 tumor-derived heterozygous mutant molecules if the ctDNA amount is 0.1%. Furthermore, sequencing this deeply would require that the investigated sample contains 60,000 unique DNA fragments covering all genomic loci. To reliably reach this number in a testing environment, one would have to analyze cfDNA from a substantial volume of blood (>20 mL). Given the challenges of performing such deep sequencing on a genome-wide scale, most approaches explored up to now have been restricted to sequencing of predefined gene panels (eg, selected DNA fragments with genes frequently mutated in the target cancers) by either capture-based or multiplex-PCR–based approaches, all of which increase the cost and effort of such methods (table 1, figure 1).9 12 26 32 34–40 Further refinement of somatic mutation-based approaches to early cancer detection are needed to make these more affordable for implementation on a population scale. Beyond individual mutations, genome-wide mutational signatures, such as C:G>A:T changes, are known to be increased in certain regions of tumors of individuals with a smoking exposure and may improve cancer detection.41–43

Another hurdle associated with somatic mutation-based approaches is determining whether the source of the alteration is from the tumor or WBCs. If somatic mutations were found in asymptomatic individuals, additional sequencing or imaging would be needed to determine the source of the mutation and its clinical relevance.26 Leal et al reported that more than half of mutations in the TP53 gene detected in the blood were derived from WBCs highlighting the challenges of pinpointing the cellular origin of plasma mutations.44 Consequently, targeted somatic mutation analysis as a standalone approach lacks the utility needed for widespread use in cancer screening of asymptomatic individuals. It comes as no surprise that approaches relying on somatic mutations are now being combined with other biomarkers.45

Cancer screening using methylation profiles

Methylation of promoter DNA sequences is a common mechanism used to regulate gene expression.46 47 Because cancer development and progression are frequently associated with aberrant patterns of DNA methylation,48 49 and PCR and next-generation sequencing (NGS)-based methods with high analytical sensitivity are available, methylation profiling has been developed as a means to detect cancer.50 51 Methylation profiling evaluates epigenetic profiles of cells that have lysed and distributed their contents into circulation. It measures the methylation status of subsets of the ~28 million CpG sites in the genome.52 These sites are not equally distributed across the genome, and many are concentrated in ~30,000 CpG islands that are located near the promoters of genes and have a critical role in regulating gene expression.52 Since the genes expressed in each cell type are different, these patterns of differential methylation can be used to identify the cell of origin of circulating cfDNA molecules.

Methylation screening approaches range from measuring a single gene to using a panel of predefined methylation sites that have been shown to be differentially methylated in cancer cells and WBCs.53–56 PCR-based approaches to detect methylation require preselection of a set of target genes and cannot be done on an unselected genome-wide basis.57 A broader enrichment for methylation sites involves hybrid capture of small regions of the genome containing a subset of CpG sites, typically using bisulfite conversion, amplification, and sequencing to determine the methylation status at each site, and application of an algorithm to determine the presence of cancer-derived DNA in the sample and the cell of origin of the cancer-derived DNA molecule (figure 1).25 58 A more comprehensive approach to methylation profiling for cancer detection uses immunoprecipitation of cell-free methylated DNA to isolate the methylated DNA regions throughout the genome and then analyses of these by NGS (table 1).24 59 More recently, methylation approaches coupled with mutation detection from the same DNA molecule have been developed.60

Cancer screening using whole-genome sequencing

Evaluation of cfDNA fragmentation patterns, known as ‘fragmentomes’, represent the most recent advancement in cfDNA-based screening modalities.61 These emerging new approaches focus on the physical characteristics of cfDNA fragments, which vary markedly across the genome in cfDNA derived from tumor cells compared with WBCs.61–63 Early observations suggested that tumor-derived cfDNA is shorter overall when compared with non-cancer cfDNA molecules.64 More recent findings showed that mutated tumor-derived DNA from different regions in the genome could be smaller or larger than wild-type non-cancer cfDNA.61 Additionally, these studies revealed that germline mutations as well as variants related to clonal hematopoiesis, which can confound mutation-based analyses, harbor cfDNA fragment size distributions that are similar to non-cancer cfDNA and therefore should not influence fragment size based approaches.44 61 65 This previously unappreciated knowledge provides a robust foundation for genome-wide fragmentome based approaches focused on early cancer detection.

The intracellular and extracellular DNA fragmentation processes associated with apoptosis and necrosis involve enzymes including caspases and endonucleases that cleave chromosomal DNA molecules into short fragments. These cleavage sites are non-random and determined in part by the association of the DNA molecules with the nucleosomal proteins, which protect the DNA from endonuclease digestion.66 On a genome-wide level, cfDNA fragmentation has been shown to reflect broader patterns of chromatin structure.66 The cellular genome is organized into two basic patterns known as open and closed chromatin that is mechanistically linked to the regulation of gene expression. The regions of open and closed chromatin are highly conserved and characteristic of the unique gene expression profile in each cell type. The three-dimensional chromatin organization is quite different in the open and closed chromatin and controls binding of the RNA transcription complex to the template DNA. This three-dimensional structure also affects the binding of other large protein molecules and complexes, such as the endonucleases and caspases, to the DNA resulting in reproducible patterns of cleavage in each cell type. Consequently, the cleavage patterns determined by the basal chromatin organization can be used to differentiate between cfDNA derived from different cell types, and between cancer and normal cells. The result is that the cfDNA fragmentome is both characteristic of the cell type/organ from which it originates, and distinct between cancer and normal cells.61

Historically, only a few approaches had leveraged genome-wide cfDNA analyses, including those to assess changes in chromosomal copy number,19 nucleosomal footprints,66 fragment lengths,67 transcription start sites68 or transcription factor binding sites.69 Genome-wide analyses of cfDNA fragmentomes led to the realization that changes in cfDNA size and coverage throughout the genome using machine learning models could be used for early detection61. Additional analyses from whole genome sequencing of cfDNA have included end motifs,70 and fragment end position aberrancy.71 These recent advances in cfDNA fragmentation biology provide orthogonal approaches to detect cancer beyond existing methods.

The majority of cfDNA fragment-based approaches rely on whole-genome next-generation sequencing data, unlike mutation analyses that mainly use targeted tumor-specific gene panels with deep coverage (table 1, figure 1).9 12 32 34 72 More recently, whole-genome sequencing to detect mutations has emerged73 but this genome-wide sequencing approach requires tumor tissue, WBCs, and plasma biomaterials to be sequenced at high coverage (~35×), making it more suitable for identification of minimal residual disease rather than early cancer detection. Variations to this method have used whole-exome tumor-tissue sequencing to identify personalized tumor-specific targets that are subsequently analyzed in plasma after therapeutic intervention.35 37 38 74 These evolving mutation-based approaches, while effective, require substantial logistics, including in obtaining both tissue and blood specimens, as well as operational costs that may limit clinical feasibility. Alternatively, emerging advancements of fragmentomics may lower such barriers to entry not only in the setting of minimal residual disease but for early detection screening (table 1).

To this end, a recent cfDNA fragmentation approach has been described that involves the extraction of cfDNA from plasma, construction of whole-genome libraries, low-coverage whole-genome sequencing, and application of a machine learning algorithm to predict the presence of cancer DNA (figure 1). This approach has been tested in a multi-cancer cohort,61 a hepatocellular carcinoma cohort,63 as well as a large lung cancer cohort.62 Studies so far have shown that fragmentation-based detection has promising sensitivity for the detection of localized early-stage cancers (stage I/II).61 62 Because fragmentation profiles can be determined through low-coverage whole-genome sequencing and there are potentially millions of fragment changes in the cfDNA originating from a cancer cell, the cfDNA fragmentation approach has potential to be widely available at low cost, satisfying the principle in that a screening test be accessible and acceptable to the target screening population (online supplemental table S2). Moreover, whole-genome fragmentation analyses do not require expensive gene-capture panels or bisulfite conversion of DNA prior to sequencing, a step during which substantial amounts of the DNA sample are lost (figure 1). In addition, all tumor-derived cfDNA is potentially assessed for the test, inherently increasing analytical sensitivity, compared with approaches that use only specific regions of the tumor DNA as input. That said, the steps involved in fragmentation analysis, including blood sample collection, cfDNA extraction, DNA sequencing, and machine-learning models, must be reproduced reliably for the test to remain robust, and it remains uncertain how much process variation can be accommodated in fragmentation analyses.

Impact of inflammatory conditions on cfDNA

Non-malignant diseases that involve the host immune system, such as chronic inflammation or autoimmunity, have demonstrated an increase in cfDNA levels in the blood.75 76 The elevated release of cfDNA into circulation could lead to a larger ‘haystack’ of cfDNA molecules that may affect the ability for a few low-quantity ctDNA molecules to be detected by mutation-based cancer tests. In fact, most cfDNA-focused mutation studies exclude individuals with inflammatory conditions, while only a few studies evaluate mutations in conditions such as chronic liver diseases or neoplasia from inflammatory bowel disease.77 78

By contrast, the use of cfDNA fragmentation-based approaches in individuals with inflammatory or autoimmune diseases have been reported for patients with liver or lung cancer.62 63 70 Jiang et al showed that fragment end motifs in healthy individuals and those with hepatitis B virus infection (with or without cirrhosis) are more similar to each other than to motifs of persons with hepatocellular carcinoma.70 In parallel, the development of an approach that assesses genome-wide fragment position through an information-weighted fraction of aberrant fragments value also observed that profiles from non-cancer individuals were similar to those from individuals harboring the hepatitis B virus with or without cirrhosis.71 Recently, Foda et al evaluated genome-wide cfDNA fragmentomes in cancer-free individuals with viral hepatitis or liver cirrhosis and observed low fragmentome scores comparable to those individuals without inflammation or cancer.63 In a study of a prospectively collected population of patients at high risk of developing lung cancer, cfDNA fragmentomes of individuals with or without autoimmune diseases were not statistically significant, and levels of circulating inflammatory markers C-reactive protein and interleukin 6 did not correlate with fragmentome-based scores in non-cancer individuals, suggesting that genome-wide fragmentation is not skewed by the presence of chronic inflammation.62 In summary, the current literature suggests that inflammatory diseases leading to increased cfDNA levels may have an impact on detection of cfDNA mutations but fragmentation approaches appear robust in this setting.

Proof-of-principle early cancer detection studies

Currently, multiple efforts are ongoing to develop liquid biopsy screening tests that effectively enable multi-cancer early detection (MCED) (table 2). Liquid biopsy early cancer screening in this setting necessitates a high specificity to reduce potentially adverse clinical intervention from false positives as well as to minimize overdiagnosis. To achieve such standards of test performances, cfDNA assay features have leveraged the biological oncogenic process such as mutations, protein biomarkers, methylation, and fragmentation features from plasma.61 79 80 The CancerSeek initial proof-of-concept approach across eight cancer types from a total of 1005 individuals introduced the potential applicability of mutations and protein biomarkers as a way to detect cancers early.79 While preliminary data from the CancerSeek approach was promising, clinical translation through a prospective interventional study known as DETECT-A (Detecting cancers Earlier Through Elective mutation-based blood Collection and Testing) resulted in a sensitivity of 27% for all cancer types at a specificity of 98.9% with a positive predictive value (PPV) of 19.4% (tables 1 and 2).26 Methylation based clinical trials such as the PATHFINDER study revealed a sensitivity of 29% at a specificity of 99.1% with a PPV of 38% (tables 1 and 2).81 Both DETECT-A and the PATHFINDER PPVs highlight the existing challenges for pan-cancer early detection tests as opposed to scenarios of high-risk screening that may benefit from a liquid biopsy prescreen. More recently, the SYMPLIFY study assessed individuals with signs and symptoms that may be indicative of cancer using the Galleri MCED test (ISRCTN10226380) (table 2).82 Interestingly, this analysis consisted of 5461 individuals with symptoms and resulted in a sensitivity of 66.3% (95% CI 61.2 to 71.1) at a specificity of 98.4% (95% CI 98.1 to 98.8) with a PPV of 75.5% (95% CI 70.5 to 80.1). It is important to note that, while the assay applied in PATHFINDER and SYMPLIFY is the same, SYMPLIFY examined patients with symptoms such that the study population is different and must be considered when evaluating performances and future applications.

To improve early cancer detection performance from these intial studies, multiple clinical trials are ongoing focused on detection of a variety of cancer types (Table 2), with the goal of achieving sensitivities in early stage disease that could clinically useful. One approach has been to integrate subsets of features including mutations, methylation, and genome-wide fragment characteristics in these analyses. Jamshidi and colleagues use a pan-feature comprised of a 507 targeted gene panel that evaluated single nucleotide variants (SNV) and white-blood cell corrected-SNVs, whole-genome methylation (30× coverage), and whole-genome sequencing (30× coverage) subcategorized to independently assess the features of copy number alterations (CNA), white-blood cell corrected CNA, fragment endpoints, fragment lengths, and allelic imbalance.80 One limitation of this study was that fragment features were not combined together and compared with methylation as was reported to an extent in previous research from the same group.83 Bruhm et al., recently combined genome-wide fragmentation features and regional mutational signatures from low-coverage whole genome sequencing to show an increased performance for cancer detection compared to either method alone.⁴¹ Orthogonal analyses as well as new approaches that reflect the underlying cfDNA biology and machine learning are still warranted to further confirm and extend such initial findings. Lastly, combining cfDNA features and emerging radiomics approaches may further improve detection especially in the context of high-risk screening.

Establishing clinical utility for early cancer detection tests

The National Cancer Institute of the USA defines clinical utility as ‘the likelihood that a test will, by prompting an intervention, result in an improved health outcome’.84 In the case of early cancer detection, the intervention that the test should inform/prompt is whether to proceed to confirmatory diagnostic testing and subsequent intervention or to continue in the future with the recommended regular screening. Because high sensitivity and high specificity are uncommon in a single test, tests are often divided into those that have high-enough sensitivity to effectively ‘rule-out’ the presence of cancer without a false-positive rate that causes substantial clinical burden and those that have high-enough specificity to ‘rule-in’ cancer at a sensitivity high enough to prove clinically useful. The ultimate and most clinically important objective of cancer screening is to reduce the likelihood that a screened individual dies of cancer. As the ultimate goal of cancer screening is to reduce cancer specific mortality, the ultimate endpoint would be cancer specific survival. As it easily can take over a decade to reliably evaluate this endpoint, as a rule, intermediate endpoints like a stage shift or detecting high-risk precursor lesions are used as alternatives. Care then must be taken that any intermediate endpoints used are representative of ultimate outcome.30 85–88 For some cancer types, like prostate and perhaps breast cancer, a shift in stage at diagnosis could rather reflect the ability of the screening test to detect indolent cancers that otherwise would not become symptomatic or impact the survival of the patient. As early cancer detection tests come of age, how do we safeguard against overdiagnosis of cancer? According to the principles of screening (online supplemental table S2), do we adequately understand the natural histories of the types of cancer likely to be detected? Stage shift per se may not translate into improved outcomes for some cancer types, and other factors, such as the availability of effective treatments, matter more. Indeed, as the principles of screening point out (online supplemental table S2), there should be ‘follow-up care that will modify the natural history and clinical pathway for the disease or condition….’ With respect to early cancer detection testing, how would we determine the magnitude of stage shift needed for each cancer type that would result in a survival benefit? Are there interventions available that could modify the natural history and clinical pathway for every type of cancer that could be detected early? If not, how should we measure the clinical utility of early detection for these less treatable cancer types? The answers to these questions are even more pertinent in the case of ctDNA screening tests that can identify multiple different types of cancer at once. Addressing these important points may require the conduct of randomized controlled trials of testing versus no testing, with endpoints that include both 5-year cancer-specific survival and overall survival, along with parallel studies to measure quality of life of those tested and untested, and cost-effectiveness of testing and no-testing strategies.

Some have argued that the endpoint for early cancer detection tests should be all-cause mortality,89 because overall mortality is what matters most to patients, it minimizes the effects of misclassification of causes of death, and showing actual lives saved is the most direct demonstration that early cancer detection saves lives. Few cancer screening trials have demonstrated significant improvements in all-cause mortality. The National Lung Screening Trial is the only single-cancer screening trial to demonstrate a significant reduction in all-cause mortality in a population with heavy smoking histories.1 With median follow-up of 6.5 years, the rate of death from any cause was 6.7% lower (95% CI 1.2 to 13.6; p=0.02) in the arm that underwent low-dose computed tomography (n=26,722) versus chest radiography (n=26,732). The multi-cancer Prostate Lung Colorectal Ovarian trial (n=154,887; median follow-up of 17 years) demonstrated a small but significant reduction in overall mortality (2%; p=0.036) for those in the screening arm versus the usual care arm, particularly for male participants.90 Because reduced overall mortality is harder to demonstrate in clinical trials, the use of such as cancer-specific mortality becomes more attractive.