Annual Review of Pathology: Mechanisms of Disease
`Detection and Diagnostic
`Utilization of Cellular and
`Cell-Free Tumor DNA
`
`Jonathan C. Dudley1 and Maximilian Diehn2
`1Ludwig Center, Department of Oncology, Johns Hopkins University School of Medicine,
`Baltimore, Maryland 21287, USA
`2Department of Radiation Oncology, Stanford Cancer Institute, and Institute for Stem Cell
`Biology and Regenerative Medicine, Stanford University School of Medicine, Stanford,
`California 94305, USA; email: diehn@stanford.edu
`
`Keywords
`cell-free DNA, cfDNA, circulating tumor DNA, ctDNA, molecular
`genetic pathology, oncology, next-generation sequencing, benign clonal
`expansions
`
`Abstract
`Because cancer is caused by an accumulation of genetic mutations, mutant
`DNA released by tumors can be used as a highly specific biomarker for can-
`cer. Although this principle was described decades ago, the advent and falling
`costs of next-generation sequencing have made the use of tumor DNA as a
`biomarker increasingly practical. This review surveys the use of cellular and
`cell-free DNA for the detection of cancer, with a focus on recent technolog-
`ical developments and applications to solid tumors. It covers (a) key princi-
`ples and technology enabling the highly sensitive detection of tumor DNA;
`(b) assessment of tumor DNA in plasma, including for genotyping, minimal
`residual disease detection, and early detection of localized cancer; (c) de-
`tection of tumor DNA in body cavity fluids, such as urine or cerebrospinal
`fluid; and (d) challenges posed to the use of tumor DNA as a biomarker by
`the phenomenon of benign clonal expansions.
`
`199
`
`Annu. Rev. Pathol. Mech. Dis. 2021. 16:199–222
`
`First published as a Review in Advance on
`November 23, 2020
`
`The Annual Review of Pathology: Mechanisms of Disease
`is online at pathol.annualreviews.org
`
`https://doi.org/10.1146/annurev-pathmechdis-
`012419-032604
`
`Copyright © 2021 by Annual Reviews.
`All rights reserved
`
`Annu. Rev. Pathol. Mech. Dis. 2021.16:199-222. Downloaded from www.annualreviews.org
`
` Access provided by 12.51.190.113 on 09/07/23. For personal use only.
`
`Personalis EX2166
`
`Personalis EX2181
`
`

`

`INTRODUCTION
`DNA is released from tumors into surrounding fluids through apoptosis, necrosis, or active se-
`cretion. DNA nucleases digest DNA between nucleosomal positions, resulting in fragmentation
`to a length of approximately 170 base pairs (1). The concept of using tumor-specific mutations
`as biomarkers for cancer was first described in 1991 through the detection of TP53 mutations in
`the urinary cells of patients with bladder cancer, with follow-up studies in 1992 describing KRAS
`mutations in the stool of patients with colorectal cancer (CRC) and, in 1994, mutations in both
`genes in the sputum of patients who developed lung cancer (2–4).
`In 1997, fetal-specific DNA sequences were identified in maternal plasma (5). Follow-up stud-
`ies demonstrated the utility of cell-free fetal DNA for the detection of chromosomal abnormali-
`ties, and it is now routinely used for the noninvasive prenatal diagnosis of trisomies and determi-
`nation of gender (6, 7). While some studies have demonstrated the possibility of using cell-free
`fetal DNA to screen for single-gene disorders, the background noise and cost of whole-exome
`or genome sequencing have made this application more challenging (8–10). Targeted testing to
`assess for predefined single-gene mutations may be more technically feasible (11). A final major
`application of cell-free DNA (cfDNA) is the detection of transplant rejection from the presence
`of increased, graft-specific DNA sequences in the recipient bloodstream (12, 13).
`This review examines the use of cfDNA as a biomarker for solid tumors, with a focus on cfDNA
`in the plasma and both cellular DNA and cfDNA in body cavity fluids. Interest in cell-free tumor
`DNA has exploded in recent years, and several excellent reviews have covered the topic in depth
`(1, 14–16). This review therefore focuses on the most significant and recent developments in this
`rapidly growing field, rather than attempting to exhaustively survey historic contributions. Par-
`ticular attention is paid to technological and conceptual milestones in detecting cell-free tumor
`DNA, the use of circulating tumor DNA (ctDNA) to detect minimal residual disease and screen
`for localized cancers, analysis of tumor DNA in body cavity fluids, and challenges posed to the
`field by recently described benign clonal expansions.
`
`CONCEPTS AND TECHNOLOGY
`Single-Locus Assays
`Initial examinations of the diagnostic potential of tumor DNA focused on single loci and employed
`PCR-based target enrichment strategies (Table 1). Mutant amplicons were detected through
`oligomer-specific hybridization, mutant allele–specific PCR, or other PCR-based strategies (2,
`3, 17, 18). More recently, the Roche Cobas EGFR real-time quantitative PCR (qPCR) assay be-
`came the first US Food and Drug Administration–approved test for tumor-specific mutations in
`cfDNA. This assay detects mutations that predict the response to tyrosine-kinase inhibitors, as
`well as resistance mutations (19). An advantage of PCR-based tests is their quick turnaround time
`and low cost. With the exception of mutant allele–specific PCR, however, they typically have a
`limit of detection above 0.5%, limiting their applicability to the assessment of advanced cancers.
`A conceptual advance in the analysis of cfDNA came with the development of digital PCR in
`1999 (20). A challenge in employing regular PCR-based methods for the detection of rare variants
`is that mutations occurring due to polymerase errors during PCR can be difficult to distinguish
`from rare variants present in the original sample. Digital PCR overcame this limitation by isolating
`the original DNA molecules prior to amplification. This was initially achieved by the dilution of
`DNA across a PCR plate at a concentration of 0.5 copies per well, ensuring that most wells only
`had one or zero starting DNA molecules (20). In current applications of digital PCR, DNA is
`typically diluted across emulsions, increasing throughput. The locus of interest is then amplified by
`
`200
`
`Dudley (cid:129) Diehn
`
`Annu. Rev. Pathol. Mech. Dis. 2021.16:199-222. Downloaded from www.annualreviews.org
`
` Access provided by 12.51.190.113 on 09/07/23. For personal use only.
`
`Personalis EX2166
`
`Personalis EX2181
`
`

`

`Table 1 Technology and applications for cell-free and cellular DNA analysis
`
`Test type
`Single-locus
`
`Genomic
`space
`1–100 bp
`
`Multilocus
`
`≥100–500 kbp
`
`Genome-wide
`
`Whole genome
`
`Techniques
`Allele-specific qPCR
`Digital PCR
`UMI PCR (e.g.,
`Safe-SeqS)
`Multiplexed PCR (e.g.,
`TAM-Seq)
`Hybrid capture (e.g.,
`CAPP-Seq)
`
`Shallow whole-genome
`sequencing (±
`bisulfite conversion)
`Genome amplification
`(e.g., REALSeq)
`
`Variants detected
`Point mutations
`Indels
`
`Point mutations
`Indels
`Copy number
`variants
`Translocations
`Aneuploidy
`Fragment end
`profiles
`Methylation profiles
`
`Example
`applications in the
`literaturea
`Actionable point
`mutation
`detection in
`advanced cancer
`Tumor genotyping
`Minimal residual
`disease detection
`
`References
`19–29
`
`29, 32–35
`
`Early cancer
`detection
`
`13, 43–45,
`47–52
`
`Abbreviations: bp, base pair; CAPP-Seq, cancer personalized profiling by deep sequencing; kbp, kilobase pair; qPCR, quantitative PCR; REALSeq, repetitive
`element aneuploidy sequencing; TAM-Seq, tagged-amplicon deep sequencing; UMI, unique molecular identifier.
`aOnly single-locus applications currently have strong evidence of clinical utility.
`
`PCR, and the presence of mutant or wild-type DNA is detected by the use of fluorescently tagged
`DNA oligomers. In a variation of digital PCR called beads, emulsion, amplification, magnetics
`(BEAMing), the DNA is amplified on oligomer-coated magnetic beads, and fluorescently tagged
`mutant or wild-type oligomers are detected by flow cytometry (21).
`Like qPCR, digital PCR can be performed quickly and at low cost. It is therefore routinely
`employed in some clinical settings for the analysis of actionable genetic variants, such as EGFR
`mutations in lung cancer or KRAS mutations in CRC (22, 23). Unlike qPCR, however, digital PCR
`has a very low limit of detection (0.01–0.1%) and has therefore also been used in some studies for
`the early detection of resistance mutations and the detection of minimal residual disease in solid
`tumors, with applications in lung and pancreatic cancers (24, 25). A disadvantage of both qPCR
`and digital PCR is that the potential mutation of interest has to be known ahead of time and
`specifically accounted for in the assay design.
`
`Molecular Barcoding
`Next-generation sequencing offers a way to screen for mutations in a broader target space. Loci are
`first amplified by PCR, and then the entire region within the amplicon can be sequenced, enabling
`the unbiased detection of any variants within the targeted space. For high-sensitivity applications,
`a unique molecular identifier (UMI) is incorporated into one of the primers in an initial round of
`PCR (26). Like digital PCR, this strategy, called Safe-SeqS in its initial description, enables one to
`distinguish variants arising due to polymerase errors during amplification from variants present
`in the original DNA molecule. Variants present in the original molecule should be present in all
`copies with the same UMI, whereas variants arising due to polymerase errors will usually only
`be present in a subset of molecules with the same UMI. An exception is if a polymerase error
`occurs in the first cycle of PCR, in which case it can be passed on to all copies with the same UMI,
`resulting in a false positive mutation.
`
`www.annualreviews.org (cid:129) Cellular and Cell-Free Tumor DNA
`
`201
`
`Annu. Rev. Pathol. Mech. Dis. 2021.16:199-222. Downloaded from www.annualreviews.org
`
` Access provided by 12.51.190.113 on 09/07/23. For personal use only.
`
`Personalis EX2166
`
`Personalis EX2181
`
`

`

`This challenge can be surmounted by attaching adapters to the double-stranded DNA that
`uniquely tag each molecule in the duplex, as well as the duplex itself. After adapter ligation, the
`strands are amplified separately, and target enrichment is typically accomplished through hy-
`bridization to oligonucleotide probes complementary to the regions of interest. Two types of error
`suppression can be performed using this approach. The first compares copies of each single strand
`of the original DNA duplex (i.e., it analyzes the strands individually). The second additionally re-
`quires a mutation to be present on both single strands of the original DNA duplex. False positive
`mutations are extremely rare with the latter approach because they would have to occur separately
`at the same spot on each strand of the duplex (27). This strategy therefore offers the lowest limit
`of detection of any available method, but it is limited by a low rate of recovery of both duplex
`strands (28, 29). Clinical studies employing this approach, therefore, have either begun with a
`massive quantity of DNA, beyond what is available from a normal blood draw, or integrated it
`with a strategy relying primarily on single-strand UMIs (29, 30).
`
`Multilocus Assays
`PCR-based assays have limited multiplexability. With qPCR and digital PCR, it is difficult to as-
`say for more than one mutation in a single reaction pool. When assaying for multiple mutations,
`therefore, the original sample is typically split across multiple separate reaction pools, a require-
`ment that may pose challenges when cfDNA is limited. Single-locus next-generation sequencing
`strategies allow multiple targets to be assessed in the same reaction pool. Without technically
`complex strategies that require massive DNA input, however, they typically cannot be used to
`assess more than 1,000–2,000 base pairs of genomic space in a single reaction pool (31, 32).
`Hybrid capture–based target enrichment strategies offer a way to overcome this limitation, en-
`abling the enrichment and sequencing of tens of thousands to millions of base pairs in a single assay.
`These strategies begin with the ligation of adapters to the original DNA molecule and have the
`advantage of preserving unique fragment ends. In combination with UMIs in the adapters, these
`unique fragment ends can be used to identify PCR copies originating from the same molecule
`and may themselves provide additional diagnostic information, as discussed below (29, 33, 34). In
`some applications, the UMIs may be coupled with identifiers that allow the original DNA duplex
`to be uniquely tagged, enabling ultrasensitive duplex sequencing when both strands are recovered
`(29).
`Hybrid capture assays that cover a large genomic space are often used in advanced cancer pa-
`tients to genotype tumors from the DNA they shed into the plasma. In some instances, assays may
`be designed to detect copy number variations within the genomic space as well as translocations,
`indels, and single-nucleotide variants, providing a comprehensive genetic portrait of a patient’s
`tumor from a noninvasive blood draw (29, 33–35).
`An additional advantage of hybrid capture assays is that they may help surmount the challenge
`posed by stochastic sampling error. When tracking a single locus at a low mutant allele fraction,
`the probability that a fragment of mutant DNA will be absent from a single blood draw by chance
`may be high due to the low concentrations of ctDNA present in most cancer patients. However,
`the more loci one tracks, the higher the probability that at least a few mutant DNA molecules
`will be present in a given blood draw. This strategy, called cancer personalized profiling by deep
`sequencing (CAPP-Seq) in its initial implementation, enables the detection of ctDNA at levels
`from about 0.1% mutant allele fraction to as low as approximately 0.0001% (1, 29, 33). This
`may be especially helpful in the setting of minimal residual disease detection, as discussed below,
`when mutations in a patient’s tumor are already known. In this context, the genomic space actually
`analyzed from a hybrid capture panel can be limited to the specific bases that were mutated in a
`
`202
`
`Dudley (cid:129) Diehn
`
`Annu. Rev. Pathol. Mech. Dis. 2021.16:199-222. Downloaded from www.annualreviews.org
`
` Access provided by 12.51.190.113 on 09/07/23. For personal use only.
`
`Personalis EX2166
`
`Personalis EX2181
`
`

`

`patient’s tumor, lowering the background noise of the panel. The identification of any one of those
`mutations may be sufficient to diagnose residual disease (36–38).
`A challenge with hybrid capture approaches is that a larger genomic space provides more op-
`portunities for artifactual mutations to arise. One strategy for reducing this problem is to first
`assess the pattern of stereotypic artifacts in a group of normal samples and informatically account
`for that background when assessing patient samples for the presence of oncogenic mutations (29,
`39, 40). An additional limitation is that hybrid capture–based workflows typically have longer
`turnaround times, since the target enrichment process often takes days (instead of hours with
`PCR), and are more expensive, since they involve sequencing a larger genomic space. Finally, de-
`spite the advantages of overcoming stochastic sampling error, the larger genomic space provided
`by hybrid capture strategies may provide diminishing additional value in some contexts (41).
`
`Genome-Wide Assays
`In the context of cancer detection, hybrid capture–based strategies typically rely on identifying
`the presence of cancer from the occurrence of known oncogenic changes within the panel space.
`However, a substantial fraction of tumors of any given cancer type may not harbor known onco-
`genic mutations, or they may harbor a small enough number of such mutations that they may be
`missed by stochastic sampling error, even with a large panel.
`Genome-wide assays can potentially address this challenge. One class of assays relies on the
`detection of chromosomal abnormalities, or aneuploidy. Aneuploidy is an early event in neopla-
`sia, present in over 90% of tumors of any given cancer type, and highly specific for cancer (42,
`43). It can be assessed either via shallow whole-genome sequencing or through PCR-based strate-
`gies. One recently developed PCR-based strategy, called repetitive element aneuploidy sequencing
`(REALSeq), amplifies short interspersed nuclear elements throughout the genome with a single
`primer pair, enabling the identification of aneuploidy from lower DNA quantities and with less
`sequencing than is possible with whole-genome approaches (44). In an initial study, REALSeq
`was able to detect 71% of cases with variant allele fractions ≥0.5% ctDNA (44). The sensitivity
`of aneuploidy analysis could potentially be improved through the preanalytic or informatic en-
`richment of DNA sequences under 150 base pairs, which have been shown in multiple studies to
`contain a higher tumor fraction (45–47).
`An alternative genome-wide strategy focuses on unique fragment end profiles. It is thought
`that cfDNA is released by apoptosis (among other mechanisms), a process involving the diges-
`tion of DNA between nucleosomes. Since the pattern of nucleosomal packaging differs between
`cancerous and normal cells, the unique fragment ends created by this digestion may be used to
`identify cancer, with potentially thousands of informative fragment ends spread throughout the
`genome (48). Additionally, since the pattern of nucleosomal packaging differs between different
`tissue types, this strategy may also be used to provide information on the tissue of origin of iden-
`tified tumors (49, 50).
`A final genome-wide strategy focuses on patterns of methylation. Since genome-wide methy-
`lation patterns vary between cancer and normal tissues and between different tissue types, this
`strategy may also be used not only to identify the presence of cancer but also to get insight into
`its tissue of origin. Methylation may be assessed through bisulfite conversion prior to sequencing
`or through targeted capture of methylated base positions (13, 51, 52).
`Genome-wide fragment end and methylation patterns have less preestablished specificity for
`cancer than do oncogenic mutations or aneuploidy. Identifying patterns specific for cancer, there-
`fore, relies on machine learning across large sample sets, a process that holds increased risk for
`overfitting and batch effects, particularly in studies without independent validation cohorts. While
`
`www.annualreviews.org (cid:129) Cellular and Cell-Free Tumor DNA
`
`203
`
`Annu. Rev. Pathol. Mech. Dis. 2021.16:199-222. Downloaded from www.annualreviews.org
`
` Access provided by 12.51.190.113 on 09/07/23. For personal use only.
`
`Personalis EX2166
`
`Personalis EX2181
`
`

`

`preliminary studies have shown the potential of these strategies (34, 41, 49, 51), independent
`follow-up studies with larger cohorts will be necessary to confirm their ability to achieve com-
`parable or superior performance relative to mutation-based strategies.
`
`Mutant Allele Enrichment
`Various strategies have been developed to enrich for tumor DNA prior to the application of the
`above methods, a process often referred to as mutant allele enrichment. One class of strategies
`takes advantage of differential melting temperatures created by mismatches between mutant and
`wild-type DNA. Coamplification at lower denaturation temperature PCR selectively amplifies
`mutant molecules by exploiting the lower denaturation temperature of heteroduplexes that form
`between mutant strands and wild-type DNA (53). Differential strand separation at critical tem-
`perature relies on multiple rounds of hybridization to biotinylated oligonucleotide probes, with
`preferential melting and amplification of mutant molecules that mismatch reference sequence
`probes (54). A separate strategy involves the hybridization of short reference sequence probes to
`the DNA followed by digestion with a double-stranded DNA nuclease that is unable to digest a
`duplex with a mismatch (55). Another class of strategies exploits the base-pair resolution speci-
`ficity of Cas9. In one study, Cas9 was used to selectively digest the wild-type allele of KRAS (56).
`In another, deactivated Cas9 was used to selectively bind a mutant allele prior to assessment by
`qPCR (57).
`These methods have demonstrated tens-fold to hundreds-fold enrichment of tumor DNA.
`However, they typically are limited to the assessment of single loci or small numbers of loci, require
`additional steps that may result in a loss of mutant molecules, and can make it harder to quantify
`the tumor proportion in the original sample. As a result, their application in clinical settings has
`been limited.
`
`TUMOR DNA IN PLASMA
`Tumor Genotyping
`Multiple studies have shown high concordance between mutations found in ctDNA and tumor
`biopsies, making it possible to genotype patient tumors from a blood draw. One study examined
`matched tumor biopsies and cfDNA from a series of 41 metastatic prostate and breast cancer
`patients with at least 10% tumor-derived cfDNA. Using exome sequencing, the authors found
`88% concordance for clonal somatic mutations and 80% concordance for copy number alter-
`ations (58). These findings are comparable to those of studies using more focused panels (33, 34,
`59). Concordance is typically higher for mutations that are clonal in the tumor than for those that
`are subclonal (34, 60). Mutations detected in cfDNA but not present in the tumor may represent
`subclonal tumor mutations not sampled in the matched tumor specimen, technical errors (as dis-
`cussed above), or benign clonal proliferations (discussed below) (61). Many of the latter variants
`can be eliminated through paired sequencing of white blood cells (59).
`Because cancer DNA is released from tumors throughout the body, it can better capture in-
`tratumor genetic heterogeneity than a single tissue biopsy. The extent of this heterogeneity at
`the time of diagnosis and its clinical importance continues to be investigated. For example, in the
`2017 TRACERx study, authors sequenced 327 regions from 100 pretreatment early-stage tumors
`from non–small cell lung cancer (NSCLC) patients. Mutations in some driver genes (EGFR, MET,
`BRAF, and TP53) were almost always clonal, but subclonal driver gene mutations with evidence
`of occurrence late in tumor evolution were also observed in 75% of patients (including mutations
`
`204
`
`Dudley (cid:129) Diehn
`
`Annu. Rev. Pathol. Mech. Dis. 2021.16:199-222. Downloaded from www.annualreviews.org
`
` Access provided by 12.51.190.113 on 09/07/23. For personal use only.
`
`Personalis EX2166
`
`Personalis EX2181
`
`

`

`in PIK3CA and NF1 and copy number alterations) (62). These results contrast with the results of
`a 2018 study of 76 pretreatment metastases from 20 patients with a variety of solid tumor types.
`The authors found that the vast majority of driver gene mutations were shared among metastases
`and that driver genes mutations that were not shared were unlikely to be functionally signifi-
`cant (63). Notably, tumor samples were sequenced to a median depth of 65× in the 2018 study
`( J. Reiter, personal communication) compared to 431× in the TRACERx study (62), potentially
`contributing to some of the differences in observations.
`One arena in which tumor genotyping from cfDNA offers an advantage is in identifying sub-
`clonal mutations conferring resistance to targeted therapy. Although these may be missed by a
`single tumor biopsy in patients with a high burden of disease, multiple studies have documented
`their emergence over the course of targeted therapy using cfDNA assays (23, 25, 64, 65). Examples
`of clinical scenarios in which targeted cfDNA genotyping is often performed include documen-
`tation of the emergence of oncogenic KRAS variants in CRC patients undergoing treatment with
`targeted anti-EGFR antibodies such as cetuximab or of the EGFR T790M variant in NSCLC
`patients undergoing treatment with anti-EGFR tyrosine kinase inhibitors (14). However, given
`the potential risk of false negatives due to low ctDNA concentration, standard clinical practice
`usually involves reflexing to tissue biopsies in patients with negative plasma genotyping results
`(66).
`In addition to detecting known resistance mutations, several studies have illustrated the poten-
`tial of serial cfDNA genotyping to discover new mechanisms of resistance. An early study used se-
`rial exome sequencing of cfDNA from six patients with advanced ovarian, breast, and lung cancers
`to document the emergence of PIK3CA mutations during treatment with paclitaxel, RB1 inacti-
`vation during treatment with cisplatin, and MED1 inactivation during treatment with tamoxifen
`and trastuzumab (67). In a more recent example, a hybrid capture panel focused on genes mu-
`tated in NSCLC was used to analyze cfDNA from 43 patients over the course of treatment with
`the third-generation EGFR inhibitor rociletinib. The authors identified novel EGFR variants and
`MET amplification as the most common resistance mechanisms, with additional recurrent vari-
`ants identified in PIK3CA, ERBB2, KRAS, and RB1 (68). A similar study in patients treated with
`the EGFR inhibitor osimertinib identified EGFR C797S mutations as a common mechanism of
`treatment resistance to this drug (69).
`
`Tumor Burden Assessment
`In addition to documenting tumor genome evolution over the course of therapy, cfDNA also
`enables assessment of tumor burden and response to therapy. Multiple studies have shown a strong
`correlation between cfDNA levels and tumor burden, and some have shown that alterations in
`levels of ctDNA (which has a half-life in the bloodstream of roughly 30 min to 2 h, depending
`on the study) provide an earlier measure of treatment response than imaging or protein assays
`(70–73).
`One early study in this area compared ctDNA, cancer antigen 15-3, and circulating tumor cells
`at sequential time points in 30 women with metastatic breast cancer undergoing chemotherapy.
`After whole-genome sequencing of the patients’ tumors, personalized assays were designed to
`monitor ctDNA levels in each patient and compared to the other biomarkers. The study found
`that ctDNA levels demonstrated better correlation with tumor burden, and over a wider range.
`It also detected treatment response earlier than the other biomarkers (74). Multiple additional
`studies, using a variety of different methods, have shown correlation between ctDNA levels, tumor
`volume as assessed by imagining, and protein biomarkers, with ctDNA outperforming alternative
`methods for tracking tumor burden (33, 73).
`
`www.annualreviews.org (cid:129) Cellular and Cell-Free Tumor DNA
`
`205
`
`Annu. Rev. Pathol. Mech. Dis. 2021.16:199-222. Downloaded from www.annualreviews.org
`
` Access provided by 12.51.190.113 on 09/07/23. For personal use only.
`
`Personalis EX2166
`
`Personalis EX2181
`
`

`

`After attempted definitive treatment for cancer with surgery or radiation, a common question
`is whether or not the patient was truly cured or if the patient has minimal residual disease. If
`the patient was not cured, he or she may benefit from further systemic therapy. This question
`can be difficult to answer by imaging, as inflammation or fibrosis can be hard to distinguish from
`recurrent tumor. Protein biomarkers, likewise, may not be available for a particular tumor type or
`may correlate poorly with patient outcome. However, ctDNA has shown significant promise in
`addressing this problem.
`Using a form of digital PCR called BEAMing, a 2008 study of 18 patients with CRC demon-
`strated that ctDNA outperformed carcinoembryonic antigen (CEA) in predicting recurrence-free
`survival following surgery. Notably, patients with undetectable ctDNA had 100% recurrence-free
`survival. The authors also demonstrated correlation between ctDNA levels and patient treatment
`course, with levels decreasing following surgery and roughly tracking with findings on imaging
`and with CEA levels (73). A 2016 study from the same group examined serial ctDNA levels from
`230 patients with stage II CRC. Using the single-locus next-generation sequencing strategy Safe-
`SeqS, the authors found that the presence of ctDNA after curative-intent surgery had a hazard
`ratio of 18 (75).
`Similar findings have been noted in other solid tumor types. A 2015 study used personalized
`digital PCR assays to detect ctDNA in patients with early-stage breast cancer following neoadju-
`vant chemotherapy. The presence of ctDNA had a hazard ratio for recurrent disease of 25, with a
`median lead time over clinical relapse of 7.9 months (76). A 2017 study using CAPP-Seq targeted
`recurrently mutated regions in NSCLC. The authors examined cfDNA from 40 patients treated
`with curative-intent radiation for stage I–III NSCLC. They detected ctDNA at the first posttreat-
`ment time point in 94% of patients who developed recurrent disease, with a median lead time over
`radiographic recurrence of 5.2 months. By using a panel covering a median of six mutations per
`patient in NSCLC, this strategy avoided the need to design personalized assays for each patient.
`Additionally, the authors found that their strategy of tracking multiple mutations per patient re-
`sulted in an increased likelihood of detecting ctDNA, likely due to the reduced susceptibility of
`this approach to stochastic sampling error (38). Additional studies have confirmed the value of
`ctDNA analysis to detect minimal residual disease in NSCLC and other tumor types (24, 36, 37,
`77–79).
`
`Tumor Detection and Classification
`Perhaps the most potentially impactful, but also the most challenging, application of ctDNA is for
`the early detection of cancer. Every available treatment modality for cancer—surgery, radiation,
`chemotherapy, immunotherapy—is more likely to lead to a cure if the tumor is detected early
`(66). But using cfDNA to screen for cancer faces two challenges: The test must have very high
`specificity and be able to provide information on the tumor’s tissue of origin (80, 81).
`Information on a tumor’s tissue of origin is needed to effectively treat it, as almost all avail-
`able treatment strategies are tailored to the tumor type. Localized tumors, in particular, are typ-
`ically treated (and can often be cured) with either surgical resection or radiation. The need for
`high specificity can be illustrated by considering the prevalence of cancer in different populations.
`Globally, the prevalence of cancer among those between 15 and 49 years old is 0.64% (82). If one
`were to test for cancer within this age group with a test that could detect every type of cancer and
`had 70% sensitivity and 98% specificity, roughly four out of every five positive results would be
`a false positive, a positive predictive value of about 20%. In actual practice, most tests can only
`detect a subset of cancers, reducing the effective prevalence in the population tested and thereby
`reducing the positive predictive value of the test. The positive predictive value of the test can be
`
`206
`
`Dudley (cid:129) Diehn
`
`Annu. Rev. Pathol. Mech. Dis. 2021.16:199-222. Downloaded from www.annualreviews.org
`
` Access provided by 12.51.190.113 on 09/07/23. For personal use only.
`
`Personalis EX2166
`
`Personalis EX2181
`
`

`

`improved by either improving the test’s specificity or testing a higher-risk population. The global
`prevalence of cancer among those aged 50–69 years old, for example, is 3.32%. Within this popu-
`lation, with the above hypothetical test, roughly one out of every two people with a positive result
`would have cancer, giving a positive predictive value of approximately 50% (82).
`One initial example of a screening strategy using cfDNA examined Epstein-Barr virus (EBV)
`DNA in the plasma as a biomarker for nasopharyngeal carcinoma, with which EBV is causally
`associated. The authors applied a single-locus qPCR-based test to 20,174 middle-aged men in
`Southeast Asia, a demographic with an elevated prevalence of the disease, and increased the speci-
`ficity of their test by requiring a second test to confirm initial positive results. With this strategy,
`the authors achieved a sensitivity of 97.1% and specificity of 98.6%. They detected nasopharyn-
`geal carcinomas at an earlier stage with superior outcomes compared to a historic cohort, albeit
`with a positive predictive value of only 11% (83). A follow-up study using hybrid capture and
`next-generation sequencing found this positive predictive value could be improved to 19.6% by
`considering the length and abundance of detected EBV molecules, even while eliminating the
`requirement for a confirmatory test (84). A further study adding consideration of methylation
`differences to the test increased the positive predictive value to 35.1% (85).
`Other studies have examined the possibility of using multilocus cfDNA assays targeting re-
`currently mutated driver genes to screen for many different solid tumor types. In