Detecting Liquid Remnants of Solid Tumors: Circulating Tumor DNA Minimal Residual Disease

Nearly two thirds of patients with solid tumors present with locoregional disease and are amenable to curative therapies (1). Surgery, radiotherapy, systemic therapy, or a combination of these approaches can achieve disease remission in the majority of these cases when using conventional measures of response, such as functional body imaging. However, in a significant subset of patients, small numbers of remnant tumor cells, termed minimal residual disease (MRD), can persist at levels below the detection threshold of imaging or physical exam and ultimately lead to disease relapse (2, 3).

Systemic therapy delivered after surgery (i.e., adjuvant therapy) or radiotherapy (i.e., consolidation therapy) has been shown to improve long-term survival in multiple cancer types (4–7), providing strong clinical evidence that eradication of MRD can improve rates of cure. However, for most solid cancers where adjuvant/consolidation therapy is currently part of the standard of care, the magnitude of benefit from adjuvant therapies is modest. This is likely in part because a significant subset (and sometimes the majority) of patients who receive adjuvant therapies are already cured by the preceding local therapy (4, 6, 8).

For several hematologic malignancies, detection of MRD via flow cytometry for tumor cells or quantitative molecular techniques for patient-specific aberrations has long been established as a poor prognostic factor following induction therapy. Accordingly, modification of therapy based on the presence of MRD has become a standard of care for acute lymphoblastic leukemia, acute promyelocytic leukemia, and chronic myelogenous leukemia (9, 10). Given such actionability of MRD in blood neoplasms, biomarkers that can similarly identify which patients with solid tumors harbor MRD could also have significant utility in personalizing adjuvant/consolidation therapy. However, until recently, approaches to detect MRD in solid cancers have lacked the sensitivity and specificity required for clinical application (11).

Recent work focused on circulating tumor DNA (ctDNA) has produced promising results, suggesting that this analyte could serve as a generalizable biomarker for MRD in solid cancers. Tumors release DNA into the blood that can be isolated from plasma collected via routine blood draws (12–14). Despite generally representing a small fraction of cell-free DNA (cfDNA) in blood plasma, ctDNA can be detected via polymerase chain reaction (PCR) or next-generation sequencing (NGS) assays targeting tumor-specific mutations, structural variants, copy-number alterations, and epigenetic features (15, 16).

Recent advances in molecular and computational biology have significantly improved the limit of detection (LOD) for ctDNA using somatic alterations (17, 18). These improvements have raised the possibility that ctDNA analysis could be used to identify patients harboring MRD following curative therapy of solid cancers and to guide the administration of adjuvant or consolidation therapies (19). Here, we review the current evidence that detection of ctDNA following definitive therapy is prognostic in solid tumors and discuss the promise and limitations of ctDNA MRD testing.

Several approaches have been utilized to identify known or common tumor mutations in plasma cfDNA samples. Digital PCR (dPCR) improves on conventional allele-specific PCR amplification by partitioning a DNA sample into a large number of smaller reactions to provide an absolute quantification that improves sensitivity (20). Digital PCR primers and probes can be designed to achieve very high specificity, and droplet-based approaches such as BEAMing (beads, emulsion, amplification, and magnetics) have maximized the number of individual DNA molecules that can be analyzed from a single sample (21). As a result, the detection limit of such digital PCR assays is limited in practice by the amount of cfDNA that can be isolated from a blood draw (22). With DNA inputs routinely achievable from patient plasma samples (∼30 ng), dPCR has been shown to have a detection limit of approximately 0.1% (23–25). However, due to differences in DNA input, sample quality, and analysis approaches, reported LODs vary substantially between studies. dPCR is very effective for tracking a small number of mutations identified from sequencing of tumor tissue or hotspot mutations with a high prevalence in the cancer of interest, such as KRAS mutations in pancreatic cancer (26). Due to complexities of multiplexing a large number of dPCR assays, the extent of multiplexing varies widely between studies and dPCR generally has inferior clinical sensitivity for MRD than highly parallel NGS methods monitoring multiple mutations (27, 28). Accordingly, dPCR is not a commonly preferred approach for solid tumor MRD detection in most contexts (29).

NGS, also known as massively parallel sequencing, has been incorporated into several different ctDNA analysis approaches, enabling interrogation of millions to billions of DNA molecules from a biological sample. One approach to achieve sufficient sensitivity to detect rare ctDNA molecules uses gene-specific PCR amplicons to amplify one or more genomic regions expected to harbor tumor-derived mutations prior to NGS. Several approaches, including Safe-SeqS, introduce unique molecular identifier (UMI) sequences during preparation of DNA libraries for sequencing to reduce technical errors (17). Personalized multiplex PCR can be utilized to monitor multiple patient-specific mutations identified from sequencing of tumor tissue (30), and NGS sequencing of these lesions in cfDNA can achieve remarkable sensitivity levels (31). A related set of approaches uses a combination of ligation- and gene-specific PCR primers to partially preserve cfDNA fragment end information (32, 33). Multiple commercial platforms have been developed based on these approaches, including Natera Inc.'s Signatera (34–36), ArcherDX's personalized cancer monitoring assay (37), and Inivata's RaDaR assay (38). Although beyond the scope of this review, numerous other PCR amplicon-based NGS methods have been used for ctDNA analysis in the past and have been reviewed in detail elsewhere (39–41).

DNA enrichment using hybrid capture with biotinylated oligonucleotides allows sequencing of larger targeted panels with better uniformity of coverage (42, 43). Capture-based NGS approaches such as CAncer Personalized Profiling by deep sequencing (CAPP-Seq) preserve cfDNA fragment size information and can also incorporate UMIs to minimize technical background (18). Capture panels can be used as “off-the-shelf” tools designed to target frequently mutated genes in one or more cancers of interest (e.g., AVENIO assay from Roche Diagnostics; ref. 44). Alternatively, personalized capture panels can be designed for each patient to enrich for patient-specific mutations identified from tumor sequencing (refs. 45, 46; Fig. 1A). Recently, we developed a novel capture-based ctDNA MRD assay called Phased variant Enrichment and Detection Sequencing (PhasED-Seq), which leverages multiple somatic mutations within individual DNA fragments to decrease both technical and biological error rates (see below) and improves ctDNA MRD detection limits down to 1 part per million, 30–100-fold lower than previous approaches (47).

In theory, application of whole-genome sequencing (WGS) to posttreatment plasma could allow detection of an even greater number of mutations than the aforementioned NGS approaches and not require custom panel design. However, due to assay background error rates, there are diminishing returns of increasing the number of tracked mutations past a certain point (see below). Therefore, recent work suggests that when combined with customized bioinformatics, direct WGS of plasma can potentially achieve similar but not superior limits of MRD detection as the other NGS approaches (48). Additionally, the amount of sequencing required for the much larger targeted genomic space makes the per sample costs of WGS significantly higher.

Although not a focus of this review, other analysis approaches based on DNA methylation or other epigenetic features reflecting chromatin state of tumor cells (termed “fragmentomics”) have been used to detect ctDNA (49). Although these emerging techniques have not been extensively evaluated in the context of MRD, epigenetic features may be complementary to somatic mutation tracking and may enable tumor genotype–naïve MRD detection. For example, the Guardant Reveal assay integrates somatic mutation and “epigenomic” (i.e., methylation and fragmentomic) approaches to detect ctDNA without prior sequencing of tumor DNA (50, 51). A recent study using this assay for detecting MRD in colorectal cancer demonstrated that incorporating epigenomic analysis improved sensitivity compared with tumor-naïve somatic alterations alone (52). However, methylation-based approaches have been reported to have LODs of approximately 0.1% (53, 54), and rigorous LOD analyses for fragmentomic-based approaches in the context of ctDNA MRD have not been published to date. Therefore, although more studies are clearly needed, it seems unlikely that these approaches will be able to match the LODs of the most sensitive tumor genotype–informed, somatic mutation–based ctDNA MRD approaches.

The LOD for an assay is the lowest quantity of an analyte that can be reliably detected (55). Unfortunately, most studies in the ctDNA field have not rigorously defined or established the LOD. To conform to clinical laboratory testing guidelines, the LOD should ideally be defined as the ctDNA concentration at which 95% of clinical samples will be called positive (56). For mutation-based NGS ctDNA MRD assays, the LOD depends on both biological and technical factors. In practice, reliable LODs have not been published for most ctDNA analysis approaches, and multiple definitions have been used, making it challenging to compare across studies. In order to be meaningful, LODs should be determined in settings that match clinical samples as closely as possible, including in the amount of cfDNA input (57).

Regardless of the detection limit, for noninvasive detection to be feasible using circulating nucleic acids in blood plasma, tumor DNA must first be released into the blood and collected in a blood draw. Although ctDNA is generally thought to be released from necrotic or apoptotic tumor cells (58–60), the exact mechanisms of ctDNA release into the bloodstream and the relative contributions of different cell death mechanisms, phagocytosis, exocytosis, active secretion, and other cellular processes have not been well characterized in human cancers (61). Across detection platforms, ctDNA levels generally correlate with tumor burden on imaging (28, 31, 46, 62, 63). Accordingly, the number of residual cells after therapy likely also correlates with ctDNA MRD levels. Prior studies have also demonstrated that ctDNA release varies within and between tumor types and histologies, with a recurring observation that squamous cell carcinomas tend to shed higher ctDNA levels than adenocarcinomas (31, 46, 64, 65). Furthermore, levels of normal cfDNA increase with tissue injury from inflammation and ischemia related to nonmalignant conditions, surgery, or even vigorous exercise (66), which can lower the allele fraction of ctDNA molecules below the LOD of an assay. Indeed, postsurgical inflammatory changes have been shown to induce a significant increase in cfDNA levels postoperatively for up to three to four weeks (67), suggesting that ctDNA MRD should not be measured immediately following surgery. Collectively and irrespective of the assay used, these biological factors can substantially affect the sensitivity of ctDNA MRD techniques for accurately detecting residual disease and predicting relapse.

Assuming that ctDNA molecules are present in a blood sample, detection of these molecules relies on their efficient profiling. Each of the ctDNA MRD assay approaches includes multiple molecular biology steps that have imperfect molecule recovery and at any of which rare mutant ctDNA molecules could potentially be lost. For instance, when interrogated by NGS profiling, mutant ctDNA molecules need to be incorporated into the sequencing library, and this library needs to be sequenced deeply enough to observe these rare molecules in the final result. Due to losses of molecules during library preparation and sequencing, NGS methods generally have cfDNA molecule recovery efficiencies ≤50% (44, 68).

Lastly, the probability of detecting ctDNA MRD is a function of the ctDNA concentration, the number of mutations tracked, and the number of unique cfDNA molecules interrogated (Fig. 1B; ref. 18). For NGS approaches, tracking multiple mutations, using more cfDNA input, and sequencing to a greater depth can improve the likelihood of identifying mutant ctDNA molecules in a given blood sample. cfDNA input is ultimately limited by the amount of plasma that can be collected and analyzed. In patients with cancer, concerns for anemia limit the amount of blood that can be collected for ctDNA analysis (69). Although greater cfDNA input improves sensitivity across different methods, increasing input cannot increase the LOD beyond an assay's background error rate.

Even if ctDNA molecules are efficiently recovered by an NGS MRD assay, their subsequent detection relies on the successful identification and quantitation by the associated bioinformatics pipeline. Within these pipelines, a key challenge is resolving the desired ctDNA biological signal from the noise or background errors arising from various technical or biological sources. In this context, there are two major components that together determine the overall background error rate of ctDNA MRD assays: (i) technical errors leading to artifactual mutations occurring ex vivo during the various molecular biology steps and (ii) bona fide somatic variants stemming from nontumor tissues that harbor mutations.

Technical errors can be introduced during sample processing from sources such as unrepaired DNA polymerase errors arising during PCR (17) and oxidative DNA damage (70, 71). Because true tumor variants cannot be resolved below the error noise floor, these errors serve to limit the lowest possible ctDNA concentration that can be detected. To address this issue, multiple strategies have been developed to decrease technical errors. These strategies include the use “barcoding” techniques that use UMIs, in silico elimination of stereotypical background artifacts (i.e., “polishing”), and inclusion of free radical scavengers during library preparation to decrease oxidative damage (17, 18, 31, 46, 68). Collectively, these error-suppression strategies can dramatically reduce the technical errors arising ex vivo during cfDNA profiling.

Biological background secondary to somatic mutations found in cfDNA but not originating from tumor cells represents a second important source of potential false-positive mutations. Through a process called age-related clonal hematopoiesis (CH), hematopoietic stem cells can acquire mutations that can be found in both circulating peripheral blood cells and cfDNA (72). When considering peripheral blood leukocytes, patients with clonal mutations above a variant allele fraction of 2% in genes that are canonically associated with hematologic malignancies, but who do not meet the criteria for diagnosis of leukemia, are considered to have clonal hematopoiesis of indeterminate potential (CHIP; ref. 73). Because the majority of cfDNA arises from hematopoietic sources (74), CH represents a major contributor to biological background in cfDNA profiling exercises, including detection of ctDNA MRD. Importantly, the prevalence of CH variants increases with patient age, broader panels, and increasing sequencing depth (Fig. 1C) and has been found to be as high as 100% in patients 60 years or older (46, 75–77).

CH represents the dominant biological source of false-positive mutations when identifying somatic mutations in cfDNA. Importantly, sequencing peripheral blood leukocytes can help to identify cfDNA mutations due to CH. In support of this notion, a recent ctDNA study of localized non–small cell lung cancer (NSCLC) found that up to 15% of TP53 mutations in the cfDNA were attributable to CH and found in matched leukocytes but not in matched tumor (78). In a separate study, deeper sequencing of matched leukocytes in patients with NSCLC found that 40.6% of TP53 mutations present in cfDNA were also detected in matched leukocytes (46). TP53 mutations in cfDNA that were also found in leukocytes displayed less evidence of the smoking mutational signature than TP53 mutations also found in matched tumor tissue, supporting their disparate biological sources.

Although sequencing peripheral blood leukocytes is helpful for distinguishing tumor-derived mutations from CH, approximately 10% of mutations found in cfDNA of noncancer controls are not found in matched leukocytes (46). This observation suggests that some CH mutations can be missed due to low prevalence of CH subclones in the peripheral blood, perhaps reflecting CH in noncirculating hematopoietic precursors. Additionally, it is possible that mutations from nonmalignant, nonhematopoietic cells can be found in cfDNA. In support of this, recent studies have demonstrated that somatic mutations can be present in diverse nonmalignant cell types, including epithelial (79, 80), endothelial (81), stromal (82), and others (83). As detailed below, incorporating sequencing of tumor tissue to identify somatic mutations that are then monitored in plasma samples can guard against these sources of biological background.

Due to the complications posed by technical and biological background mutations described above, most ctDNA MRD studies have performed “tumor genotype–informed” analyses to monitor known tumor variants in posttreatment plasma (Fig. 1D). This approach includes the genotyping of tumor tissue to identify mutations that are then tracked in plasma, which lowers the risk of false positives due to technical and biological background sources of error. Additionally, such a tumor genotype–informed approach limits the number of genomic positions interrogated in cfDNA, therefore decreasing multiple hypothesis testing. Accordingly, the tumor genotype–informed approach can be less demanding for blood sample volumes, because fewer unique mutant cfDNA molecules are required for detecting ctDNA as compared with a genotype-naïve approach (Fig. 1E).

Several commercially available ctDNA platforms support tumor genotype–informed MRD detection, including assays for solid tumors from Natera, Roche Diagnostics, Invitae (ArcherDx), Inivata, and Foresight Diagnostics, among others. Notably, these platforms are technically distinct from liquid biopsy panels such as Guardant360 (https://www.accessdata.fda.gov/cdrh_docs/pdf20/P200010A.pdf) and FoundationOne Liquid CDx (http://www.accessdata.fda.gov/cdrh_docs/pdf19/P190032A.pdf) that have been developed for noninvasive genotyping in the setting of advanced disease (84). Although such noninvasive genotyping assays are very useful for identifying actionable tumor mutations in patients with metastatic disease (85), a recent analysis by the SEQC2 Working Group led by the FDA found that detection of variants was less reliable below an allele fraction of 0.5% (86). Therefore, currently available assays designed for the primary purpose of noninvasive genotyping in advanced disease are not optimal for ctDNA MRD detection. This is mainly due to the low ctDNA allele fractions typically observed following definitive treatment of localized solid cancers in the absence of radiographic disease burden as well as the confounding effects of CH. In contrast, tumor genotype–informed MRD detection approaches can attain LODs of ≤0.01%, which makes them preferrable for detection of minute amounts of MRD (refs. 18, 37, 47, 87; Fig. 1F).

In ctDNA MRD studies focused on solid tumors, two main types of analysis have generally been reported: (i) MRD landmark analysis and (ii) surveillance analysis (Fig. 2A). Although these two types of analysis are related, their distinct features are relevant for the clinical application of ctDNA MRD. MRD landmark analysis determines the ctDNA status of patients at a single, prespecified time point, which is typically shortly after completion of standard-of-care treatment (e.g., surgery and radiotherapy). In contrast, ctDNA surveillance analysis involves evaluating longitudinal blood draws at multiple time points during follow-up, with ctDNA status determined by whether any blood draw (irrespective of time point) is positive.

From a clinical perspective, determining MRD status at an early posttreatment landmark time is attractive because it would facilitate immediate decision-making about adjuvant or consolidation therapies, while minimizing the costs from testing serial blood samples. However, one could also envision performing repeat MRD surveillance over time and initiating adjuvant or consolidation therapy at the first positive blood draw. This approach has the potential advantage of decreasing false negatives. For either the landmark or surveillance approaches to offer the potential of meaningful improvement over current standard of care, MRD must be detectable with a significant lead time over conventional imaging. Given the differences in these two approaches, the type of analysis must be considered when comparing across studies and techniques.

More than 20 studies to date have demonstrated the prognostic value of ctDNA MRD detection in multiple solid cancers with a variety of technical approaches (Table 1). We attempted to summarize the performance of current ctDNA analysis approaches from some of these seminal studies, which we selected from the literature based on inclusion of at least 10 patients with ctDNA analysis performed after completion of curative-intent therapy for nonmetastatic solid tumors (Fig. 2B). Studies and/or patients were excluded if adjuvant therapy was given after ctDNA analysis, but this information was not available for every study. It is important to note that different methodologies for ctDNA MRD analysis have not been tested side-by-side on the same set of samples, and these analyses would be challenging due to the large amounts of plasma that would be required. Additionally, the same method may perform differently across laboratories or in samples with different quality. Our goal in this review is to synthesize the published data rather than to rigorously compare diverse methods.

Despite variable study designs, different tumor types, and inconsistent study endpoints, detection of ctDNA MRD at the posttreatment landmark was consistently associated with inferior prognosis (Fig. 2C). Hazard ratios for progression-free survival or freedom from progression in ctDNA MRD positive patients ranged from 3.5 to 43.3, demonstrating the strong prognostic value of ctDNA MRD analysis. To summarize results in more detail, for each study we calculated the clinical sensitivity for detection of residual disease (i.e., the percentage of patients who recurred who were ctDNA positive after therapy), as well as the clinical specificity (i.e., the percentage of patients without recurrence who were ctDNA negative). Studies where clinical sensitivity and specificity could not be determined from the reported results and published supplementary materials could not be included.

To compare across studies using an MRD landmark analysis, we considered the first blood draw after completion of all therapy. Of note, blood draw timing was heterogeneous across studies, and the majority did not report the length of time between completing therapy and ctDNA analysis for individual samples. Given that limitation, at the MRD landmark, clinical specificity was greater than 90% in the majority of studies. However, clinical sensitivity was more modest and ctDNA was detected in only about half of subjects who ultimately relapsed (median 56%; range, 21–100%; Fig. 2D). Therefore, patients who are ctDNA positive by landmark analysis appear to have a very high likelihood of relapse (i.e., high positive-predictive value). However, because many patients who ultimately recur are missed by current approaches, there is significant room for improving both clinical sensitivity and the negative-predictive value of ctDNA MRD. Despite this limitation, freedom from progression based on ctDNA MRD landmark analysis was strongly significant when combining data across studies (HR, 8.9; 95% CI, 4.0–19.9; Fig. 2E).

For summarizing surveillance analysis across studies, we considered patients as positive for MRD if ctDNA was detected in any longitudinal sample obtained prior to relapse on imaging. As compared with landmark analysis, extending ctDNA testing to longitudinal samples with surveillance analysis substantially improved sensitivity, approaching 100% in most studies (median 89%; range, 67%–100%; Fig. 2F). Importantly, specificity remained high despite analyzing multiple samples. Combining data across studies and compared with ctDNA MRD landmark analysis, ctDNA surveillance analysis improved stratification of freedom from progression (HR, 23.6; 95% CI, 13.0–42.9; Fig. 2G).

Summarizing across studies, residual disease was identified at a median of 5.8 months earlier using ctDNA MRD than by standard-of-care radiologic imaging. This diagnostic lead time could provide a substantial head start for initiating adjuvant or consolidation therapies and potentially lead to improved outcomes. When comparing between ctDNA analysis approaches, surveillance analysis counterintuitively had an improved lead time over the landmark approach (Fig. 3A). Furthermore, more than half of patients who ultimately developed recurrence had false-negative MRD tests at the posttreatment landmark immediately after completing therapy (Fig. 3B). Of note, it is possible that this result is affected by guarantee-time bias, because complete raw data are not available for most of the studies included in these analyses (88). This caveat aside, the apparently longer lead time in the surveillance analyses is most likely due to patients with initially false-negative MRD results having lower microscopic disease burden and/or more indolent tumors with lower ctDNA shedding that takes longer to manifest radiologically. Indeed, among patients who ultimately progressed, freedom from progression from the start of treatment was significantly longer in patients with ctDNA not detected in the first posttreatment sample compared with patients who had ctDNA detected at that time point (Fig. 3C). Thus, there is an opportunity for MRD assays with higher sensitivity to detect residual disease in patients with the lowest disease burden immediately after therapy. This would provide the longest window of opportunity before radiologic progression in which to intervene and potentially improve outcomes. Additionally, decreasing the number of false-negative MRD results immediately after local therapy could facilitate sparing patients who are already cured from unnecessary adjuvant/consolidation therapy.

As discussed above, false negatives at the MRD landmark could result from elevated normal cfDNA after surgery lowering the allele fraction of ctDNA molecules or few residual tumor cells. Consistent with the potential adverse impact of tissue injury associated with treatment and/or a progressive increase in the amount of residual disease over time, clinical sensitivity increased with a larger gap between local therapy and ctDNA analysis (Fig. 3D). Additionally, certain clinical scenarios consisting of low burden recurrence may be particularly challenging for ctDNA MRD detection. For example, isolated local recurrences after radiotherapy have been recurrently associated with false-negative MRD testing in lung, esophageal, and rectal cancers (62, 65, 89).

To help determine the LOD required for ctDNA MRD landmark analysis, we analyzed the mean ctDNA allele fraction in the first posttreatment sample for patients who developed recurrent disease in published data (Fig. 3E). The median ctDNA allele fraction post-therapy ranged from not detected to 1.6% across the studies. Across studies and tumor types, the majority of patients experiencing recurrence had ctDNA allele fractions less than 0.01% at the MRD landmark (Fig. 3F). This suggests that in order to minimize false negatives in most clinical scenarios, ctDNA MRD assay will require LODs substantially below 0.01%.

Because cfDNA input directly affects the LOD for ctDNA analysis, we also examined the cfDNA concentration at the first blood draw after completing local therapy across studies that provided this information (Fig. 3G; refs. 28, 34, 62, 90, 91). The median plasma cfDNA concentration ranged from 3.0 to 10.8 ng/mL across studies. Combining the data from these studies, the overall median cfDNA concentration was 4.3 ng/mL. Because one haploid human genome has a mass of ∼3.3 pg, this is equivalent to ∼1,303 haploid genomes per mL of plasma.

The ultimate goal of ctDNA MRD analysis is to allow personalization of adjuvant/consolidation therapy to increase the probability of cure and/or reduce toxicity. However, there is limited evidence that treatment escalation or deescalation based on ctDNA MRD status can improve patient outcomes. A growing number of published cases have suggested that adjuvant chemotherapy can lower ctDNA levels in patients with MRD following definitive therapy (31, 36, 91). In the largest study, 88 patients with stage III colon cancer underwent ctDNA testing with Safe-SeqS postoperatively and following adjuvant chemotherapy (92). Among 18 patients with evidence of ctDNA MRD positivity after surgery, 9 patients converted to ctDNA MRD negative after chemotherapy, and 6 of these patients remained disease-free at last follow-up. This suggests that ctDNA MRD can be used to track response to adjuvant systemic therapy.

A critical open question is whether adjuvant/consolidation therapy can improve clinical outcomes in patients with ctDNA MRD. Given the significant false-negative rates of ctDNA MRD landmark analysis in most studies, concerns have been raised that ctDNA MRD might mainly detect patients with the highest residual disease burden who might not be able to benefit from early initiation of additional therapy. To address this question, we recently leveraged a change in the standard of care for locally advanced NSCLC to compare outcomes based on ctDNA MRD status using CAPP-seq in patients treated with and without consolidation immunotherapy (90). Patients who were ctDNA MRD positive (but not patients who were ctDNA MRD negative) after chemoradiation and who received consolidation immunotherapy had significantly better freedom from progression than patients treated with chemoradiation therapy alone. This is the first evidence we are aware of that additional therapy can not only reduce ctDNA levels but may improve clinical outcomes in patients who are ctDNA MRD positive. In further support of the idea that adjuvant/consolidation immunotherapy can improve outcomes in ctDNA MRD–positive patients, a recent exploratory analysis of the IMvigor010 study evaluating adjuvant atezolizumab in muscle-invasive urothelial cancer demonstrated that although there was no overall survival benefit in the intent-to-treat population, adjuvant atezolizumab improved overall survival in ctDNA-positive patients (93). These two studies suggest that personalized therapy based on ctDNA MRD status could result in improved patient outcomes.

Lastly, it is possible that new mutations detected at the time of MRD analysis but that were absent pretreatment could inform therapy. Evidence from chronic myelogenous leukemia and acute lymphoblastic leukemia where molecular monitoring has become standard of care (94–96) suggests mutations identified during MRD monitoring could be used to personalize additional therapy. In solid cancers, the identification of resistant or emergent clones with actionable mutations via ctDNA analysis could potentially similarly guide the selection of adjuvant therapy (62). Indeed, noninvasive genotyping has been shown to identify actionable mutations that predict response to therapy in solid cancers (85, 97, 98). The main limitations of this approach are the low ctDNA allele fractions usually observed after definitive therapy and the relatively low detection limits of tumor-naïve ctDNA analyses, which will likely hinder the identification of clinically actionable variants in most settings.

Although many retrospective studies have demonstrated the strong prognostic power of ctDNA MRD over other disease risk factors, the clinical utility added by such assays in solid tumors remains to be determined. For example, in contrast to the established impact of MRD for outcomes in leukemias, specific evidence for the clinical benefit of ctDNA MRD analysis for personalizing adjuvant/consolidation therapy in solid tumors is scant. To firmly establish the clinical utility of ctDNA MRD for treatment personalization, prospective clinical trials will be critical. Auspiciously, ctDNA analysis has opened the door to several novel clinical trial designs in which treatment decisions are based on ctDNA MRD status (Fig. 4A; ref. 99). As of April 2021, at least 19 interventional trials are under way exploring the ability of ctDNA MRD to guide adjuvant/consolidation therapy in a variety of solid tumors using diverse randomized and nonrandomized schemas (Table 2).

The most straightforward trial design consists of treating patients who are ctDNA positive after completing the current standard of care (be it local therapy alone or a combination of local and systemic therapy) with additional systemic therapy. This trial design is attractive because it does not withhold standard-of-care treatment, and it can be implemented in solid cancers with and without a proven benefit for adjuvant therapy. Furthermore, such trials leverage the generally high positive-predictive value of ctDNA MRD assays. ctDNA-negative patients can be followed on study with active surveillance or excluded from the study. Several clinical trials are currently testing the utility of this approach in colorectal cancers (NCT04089631, NCT04120701, NCT04068103, NCT04457297, NCT03803553, EUCTR2018-000070-30-DK, and NTR6455), esophagogastric cancers (NCT04510285), NSCLC (NCT04385368, NCT04585477, and NCT04625699), triple-negative breast cancers (NCT03145961), muscle-invasive bladder cancers (NCT04660344), and solid tumors with high microsatellite instability levels (NCT03832569).

In cancers with a proven benefit for adjuvant/consolidation therapy, ctDNA testing can potentially be used to escalate or deescalate the standard-of-care adjuvant/consolidation therapy. Such strategies are currently being tested in colorectal cancer (ACTRN12617001566325) and triple-negative breast cancer (NCT04849364). Alternatively, ctDNA testing could be used to measure the response to adjuvant/consolidation therapy and escalate treatment if standard of care is not effective, as is currently being tested in NSCLC (NCT04585490). These approaches can also be combined into composite trials, as is currently being done in stage II–III colon cancers (NCT04259944). Although intensification of adjuvant therapy has often failed to improve outcomes in unselected patients when rates of cure are high with local therapy (100, 101), escalation of adjuvant therapy can have a dramatic effect on patients at high risk of developing metastatic disease (102). As a result, escalating adjuvant therapy could improve outcomes when applied to patients with ctDNA MRD who are at very high risk of relapse.

Another ctDNA MRD–driven personalization strategy that is currently being explored in low-risk stage III colon cancer (JPRN-jRCT1031200006) is the use of ctDNA to guide the omission of adjuvant/consolidation therapy in patients who are MRD negative. Such trials will need to carefully consider the false-negative rate of the ctDNA MRD assays being used to ensure that it is sufficiently low to justify such an approach. Surveillance ctDNA testing could also be incorporated into this type of trial to trigger salvage systemic therapy if patients become ctDNA positive during follow-up. For all of these studies, patients would ideally be randomized to the standard of care or a ctDNA-guided approach. Concerns have been raised by multiple stakeholders, including both patients and providers, that the high positive-predictive value of ctDNA MRD detection for recurrence makes equipoise challenging for randomizing ctDNA MRD–positive patients to no further treatment. One approach that is being used to address this issue is randomizing patients prior to ctDNA analysis, with patients on the control arm having plasma samples banked for retrospective analysis and therefore not receiving real-time ctDNA MRD results (e.g., NCT04068103).

Although the majority of ongoing interventional studies focus on traditional clinical outcomes such as recurrence-free survival (Fig. 4B), clinical trials can also be designed to use ctDNA changes as a primary or secondary outcome measure (Fig. 4C). Several trials have implemented ctDNA clearance or decrease after starting adjuvant/consolidation therapy as a disease endpoint (e.g., NCT03803553, NCT04068103, NCT04585477, NCT04585490, NCT03145961, NCT04510285, and NCT03832569). If demonstrated to correlate with traditional patient outcomes, ctDNA changes could eventually serve as surrogate for clinical benefit. Because ctDNA changes can be assessed substantially earlier than traditional outcome measures, this approach could help accelerate testing of new adjuvant/consolidation therapies.

The technical and biological aspects of ctDNA MRD testing described above are important considerations for the success of such clinical trials. The ability to improve outcomes with ctDNA-guided therapy will likely depend on the amount of residual disease that can be identified and the lead time prior to clinical recurrence. As a result, it will likely be critical to utilize ctDNA MRD assays with maximal sensitivity. Because ctDNA shedding varies between patients, it may be reasonable to exclude patients without ctDNA detected prior to treatment from clinical trials. However, these patients have been shown to have low risks of recurrence in some contexts, suggesting they might benefit from treatment deintensification analogously to patients who become ctDNA negative after initial therapy (46, 53). Clinical trials will also need to be powered sufficiently to account for only a subset of patients being either ctDNA MRD positive or negative. Furthermore, biomarker analysis must be completed in a timely manner to inform therapeutic decisions because delaying adjuvant therapy may lead to worse outcomes (103, 104). To allow intervention as early as possible, studies investigating whether MRD landmark analysis can be performed closer to the end of definitive therapy or even during standard-of-care treatment in the case of radiotherapy and systemic therapy are warranted.

A key barrier to the clinical implementation of ctDNA MRD testing in the United States will be reimbursement by payers. Based on prospective nonrandomized evidence (36), the first local coverage determination (LCD) was recently finalized by the Centers for Medicare and Medicaid Services to provide coverage for ctDNA MRD testing. The initial LCD focuses on the application of Natera's Signatera MRD test to patients with colorectal cancers. However, a broader draft LCD was recently released to enable coverage of ctDNA MRD across tumor types. The draft LCD outlines the data required for ctDNA tests to be covered by Medicare, including: (i) the ctDNA test identifies recurrence prior to clinical exams or radiographic imaging with similar sensitivity and specificity to radiology, (ii) the test must identify recurrence or progression with considerably more accuracy than other established forms of surveillance, (iii) the test must have comparable performance to other covered ctDNA MRD tests, and (iv) the test must complete a technical assessment that confirms its analytic and clinical validity. Notably absent is a requirement for randomized prospective data demonstrating improved outcomes. Thus, it is likely that an increasing number of ctDNA MRD assays for a variety of tumor types will be covered by Medicare in the near future.

Multiple studies have demonstrated that detection of ctDNA MRD following definitive therapy for solid cancers is strongly prognostic and has extremely high positive-predictive value for risk of recurrence. Due to the low allele fractions at posttreatment time points, optimizing cfDNA recovery, tracking multiple tumor genotype–informed mutations, and minimizing the effects of technical and biological background are critical to achieve high sensitivity for identifying patients at risk of progressive disease. Several interventional clinical trials testing the utility of ctDNA MRD for personalization of adjuvant/consolidation therapy are now under way, and it is likely that multiple ctDNA MRD assays could soon be used in routine clinical care. Therefore, we appear to be at the cusp of a revolution in personalized management of adjuvant/consolidation therapies in solid tumors.

E.J. Moding reports grants from Conquer Cancer and My Blue Dots during the conduct of the study; personal fees from DeciBio outside the submitted work. B.Y. Nabet reports grants from American Cancer Society outside the submitted work; in addition, B.Y. Nabet has a patent for immunmodulatory RNA pending; and is currently an employee and stockholder at Roche/Genentech and has patent filings related to immunomodulatory RNA and cancer biomarkers. A.A. Alizadeh reports personal fees and other support from Roche, Gilead, and Chugai, other support from Genentech, Celgene, Janssen, FortySeven, CiberMed Inc., Foresight Diagnostics, and Pharmacyclics, and personal fees and other support from grants from BMS during the conduct of the study; in addition, A.A. Alizadeh has a patent 20210214437 issued, licensed, and with royalties paid from FortySeven, a patent 20210172022 pending and licensed to Foresight, a patent 20210033608 pending to MARIA, a patent 10167514 issued and licensed to CiberMed, a patent 10633450 issued, a patent 20190338364 pending to CiberMed, a patent 9605320 issued to Idiotype Vaccines, and a patent 20140296081 issued, licensed, and with royalties paid from Roche. M. Diehn reports personal fees from Roche Sequencing Solutions, grants and personal fees from AstraZeneca, Illumina, Genentech, personal fees from Novartis, Gritstone Oncology, BioNTech, and Boehringer Ingelheim, grants from Varian Medical Systems, and other support from CiberMed Foresight Diagnostics outside the submitted work; in addition, M. Diehn has a patent for ctDNA detection issued and licensed to Roche and patents for ctDNA detection pending and licensed to Foresight Diagnostics.

E.J. Moding is supported by a GO2 Foundation for Lung Cancer Conquer Cancer Young Investigator Award and the My Blue Dots organization. B.Y. Nabet is a Stanford Cancer Systems Biology Scholar and supported by the NIH (5R25CA180993) and by the Postdoctoral Research Fellowship (134031-PF-19-164-01-TBG) from the American Cancer Society. A.A. Alizadeh is a Scholar of the Leukemia and Lymphoma Society and supported by the NCI (R01CA257655, R01CA233975, and R01CA188298) and the Virginia and D.K. Ludwig Fund for Cancer Research. M. Diehn is supported by the NCI (R01CA188298, R01CA233975, R01CA244526, and R01CA254179), Stand Up To Cancer, the Tobacco Related Disease Research Program, the Virginia and D.K. Ludwig Fund for Cancer Research, and the CRK Faculty Scholar Fund. The schematics in Figs. 1, 2, and 4 were created with BioRender.com.