Preclinical circulating tumor DNA shedding duration and prognostic implications of modeling 3669 patients with cancer in the American Cancer Society Cancer Prevention Study‐3 and Circulating Cell‐Free Genome Atlas Substudy 3

Abstract

Introduction

Previous studies have estimated the mean sojourn and dwell times within stages for commonly screened cancer types. However, little is known about the preclinical detection window of circulating tumor DNA (ctDNA) (ie, ctDNA positivity), which is important for understanding multicancer early detection.

Methods

The duration of preclinical detectability and prognostic value of ctDNA detection was estimated from patients with cancer in two biobank studies: CPS‐3, where cancer was diagnosed (n = 1064) within 3 years of a prior blood draw (2006–2013), and CCGA3 (NCT02889978), with a blood draw (2016–2019) concurrent with clinical diagnosis (n = 2604). To infer these quantities, Bayesian models were used for detection rates as a function of time, as well as to infer prognostic effects.

Results

Median [credible interval] sojourn times were 0.75 [0.47, 1.30], 0.89 [0.61, 1.33], 1.2 [0.84, 1.67] years for cancers diagnosed at local, regional, and distant stage, respectively, and ranged by type from pancreas, 0.49 [0.26, 0.88] to lymphoma, 2.45 [1.14, 4.87]. The extrapolated effect of ctDNA positivity at clinical diagnosis versus negative in CPS‐3 cancer cases was a relative hazard ratio of 1.98 [1.08‐4.22] for mortality.

Conclusions

These results provide estimates for average ctDNA detectable sojourn time in tumors across multiple cancer sites and stages and can inform the design of future screening studies for multicancer early detection.

Short abstract

The duration of preclinical detectability of circulating tumor DNA is estimated across multiple cancer sites and stages. These estimates can inform the design of future screening studies for multicancer early detection.

INTRODUCTION

Cancer is a progressive disease with outcomes that worsen if diagnosed at later stages. The potential to detect cancer earlier, in the preclinical phase of development before symptoms occur, is enabled by recent advances in the measurement of circulating tumor DNA (ctDNA) shed by cancers into the blood. However, the shedding dynamics of ctDNA before clinical diagnosis are not known for most cancers. The presence of detectable quantities of ctDNA has previously been associated with relatively aggressive cancers, ¹ , ² , ³ suggesting that these tumors may develop rapidly before clinical diagnosis. Understanding the duration (sojourn time) of the preclinical detection window before clinical diagnosis, either by symptomatic presentation or other means, in which early detection can occur (Figure 1), will help quantify the opportunity for early detection and inform screening interval.

FIGURE 1.

Early detection of cancer can occur in the preclinical detectable window. Cancer signal detectability starts at the limit of detection and offers an opportunity for earlier cancer detection within the preclinical detectable window, defined as from the time a detectable cancer signal appears to when a clinical diagnosis is made. This interval is sometimes referred to as the sojourn time.

Previous studies have estimated the mean sojourn time and dwell time within stages for commonly screened cancer types. ⁴ , ⁵ , ⁶ However, little is known about the preclinical detection window of ctDNA (ctDNA positivity). Assessing the ctDNA status in blood samples collected at various lead times (i.e., the interval between blood draw and clinical diagnosis) provides the information necessary to characterize timescales of preclinical cancer development as measured by ctDNA.

To undertake this analysis, we used a subset of individuals from two major studies: the American Cancer Society (ACS) Cancer Prevention Study 3 ⁷ (CPS‐3, n = 1064) and the Circulating Cancer Genome Atlas third substudy ⁸ (CCGA3; NCT02889978, n = 2605). CPS‐3 is a prospective cancer epidemiology cohort study with sufficient blood banked to apply a commercially available multicancer early detection (MCED) test evaluating cancer‐specific DNA methylation patterns in ctDNA. Individuals in this cohort were subsequently linked to cancer registries to ascertain clinically diagnosed cancers that occurred after blood draw; those individuals were followed for long‐term mortality outcomes up to 9 years after clinical diagnosis. CCGA3 provides a reference case‐control study in which the same MCED test was applied among cases where blood was drawn at clinical diagnosis.

We modeled the sojourn time for ctDNA detection across cancer types and stages by combining results of a commercially available MCED test applied to individuals drawn from both studies—one with a blood draw at various preclinical times, which provides information about change over time, and one with a blood draw at clinical diagnosis, which provides information about the final detection status. In addition, we examined the prognostic power of preclinical ctDNA detection within this window. These estimates of sojourn time and prognosis provide new insights into potential implementations of MCED tests for cancer screening.

MATERIALS AND METHODS

Model overview

We modeled preclinical sojourn time using data from two large biobank studies for which results were available from a commercially available MCED test that uses a targeted methylation assay with a locked classifier for cancer signal detection. ⁸ , ⁹

Model input populations

We used data from two large biobank studies. Individuals from the CPS‐3 study had a blood draw at enrollment (occurring between 2006 and 2013), covering approximately 300,000 participants between the ages of 30 and 65 years. ⁷ , ¹⁰ Individuals in this cohort were linked with cancer registries to ascertain clinically diagnosed cancer after the blood draw and to follow those individuals for long‐term mortality outcomes evaluated up to 9 years after blood draw. This relatively healthy, young population (50% aged <50 years; ∼70% nonsmokers; 75% female) had a lower cancer incidence (∼0.32%) than the US populations covered by the 17 Surveillance, Epidemiology, and End Results (SEER) registries from 2006 to 2015, ¹¹ matched for age and sex (0.46%). The analysis in this paper was restricted to individuals from a descriptive study with defined, stageable, single cancer types and valid MCED test data (n = 1064) in whom cancer was diagnosed within 3 years of blood draw (Supplementary Table 1 for filtering criteria). Stage at diagnosis was classified according to SEER Summary Stage, which characterizes invasive cancers as localized, regional, or distant. ¹²

CCGA3 (NCT02889978) is a case‐control study in which blood draw (occurring between 2016 and 2019) took place soon after clinical diagnosis (n = 2604). Demographic and cancer characteristics of participants in this case‐control study have been previously described. ⁸ The validation sample for this test contained individuals ranging from age 21 to 85 years, ⁹ although the expected population for screening is typically age 50 years and older. Participants with cancer types and stages matching the sampled subset of the CPS‐3 cohort were used in the joint model. Stage at diagnosis was collected according to American Joint Committee on Cancer, 6th edition, staging, which was approximated into the SEER staging system as localized, regional, or distant stage using a simple rule (localized = Stage I, regional = Stage II + III, distant = Stage IV).

Blood processing and storage

In CCGA3, blood samples were collected in Streck tubes and stored frozen for less than 5 years. In CPS‐3, blood samples were collected in EDTA tubes and stored frozen for 9 to 15 years between collection and assay. Further details are available in the descriptive study protocol. ¹³

Observed detectability

Both CCGA3 and CPS‐3 samples were analyzed using the same laboratory assay and locked classifier for cancer signal detection. Results used for this study were a cancer signal detected or not detected.

Modeling detectability

We constructed a model for detectability including the following factors (Figure 2): study, including any experimental effects resulting from differences in sample processing and storage; inherent ctDNA detectability (probability of shedding detectable quantities of ctDNA) at clinical diagnosis stratified by cancer type and stage; and effect of lead time by cancer type and stage.

FIGURE 2.

Modeling changes in cancer detection conditional on time between blood draw and detection. (A) Cancers may change state from not shedding detectable amounts of ctDNA (negative) to shedding detectable amounts of ctDNA (positive) as cancer progresses and are not expected to revert to undetectable levels once detectable. In the CPS‐3 study, we were unable to determine the detectable ctDNA status at clinical diagnosis directly. (B) Computing the probability of detection at blood draw compared to clinical detection (upper) depends on the duration between blood draw and diagnosis. The earlier the blood draw, the more likely the blood draw is to be ctDNA negative when compared to a hypothetical blood draw at clinical diagnosis. The rate of exponential change of detection depends on the cancer type and stage (lower). CPS‐3 indicates Cancer Prevention Study 3; ctDNA, circulating tumor DNA.

To account for potential sources of variation in detection related to sample collection as described previously, a factor on the logit scale was included in the model for any systematic shift in detection between the two studies.

Cancer type and stage are known to affect the probability and degree of a cancer with detectable amounts of ctDNA at clinical diagnosis and therefore the detectability of such samples.⁸ Because CCGA3 contained varying numbers of samples for each cancer type and stage, the model accounted for the stochastic uncertainty in the observed values by including a latent variable on the logit scale representing this probability for each cancer type and stage.

The effect of lead time was modeled by an exponential distribution representing the probability that the blood draw fell within an exponentially distributed window size during which ctDNA was detectable. The timescale for this exponential was modeled to vary by cancer type and separately by stage (Supplementary Methods). ¹⁴

Predicting detectability and changes over time

We used a published model of the natural history of cancer to predict how detection rates would change over lead time. ¹⁵ We modeled the proportion of interval cancers that would occur for a given screening interval from 3 months to 10 years spaced at 3‐month intervals. By definition, interval cancers are those that are not detected by a blood draw at the start of the screening interval. Taking the variations between differing intervals of screening provides an estimate of the detection rate at a given lead time between blood draw and clinical diagnosis, given a particular set of natural history parameters. Predictions for this model were compared to the observed detection rates to evaluate whether the outputs of the natural history model resembled the empirical data.

Modeling survival data

To evaluate the clinical significance of ctDNA positivity at blood draw, we did a simple hazard analysis of the difference in survival between ctDNA‐negative and ctDNA‐positive cases. This measures the potential for ctDNA positivity to find individuals who may benefit from early detection. This does not measure a causal effect of ctDNA positivity, as the cancer spectrum (cancer type and stage distribution) and lead time vary widely between ctDNA‐positive and ctDNA‐negative cancer populations.

We also modeled the association between ctDNA status at clinical diagnosis and long‐term prognosis (specifically, cancer‐specific survival [CSS]), with up to 9 years of follow‐up for mortality in selected individuals after blood draw in CPS‐3. Multiple factors are known to affect long‐term prognosis, including age, sex, cancer type, and cancer stage. These interact in complicated ways with ctDNA positivity because the cancer spectrum changes by stage, as does the expected frequency of ctDNA positivity. We accounted for these factors by using the expected CSS for a given cancer type and stage (matched for age and sex) as a reference survival curve, offset by the lead time between blood draw and clinical diagnosis, during which individuals are guaranteed to survive. Between blood draw and clinical diagnosis, the cancer can evolve from ctDNA not detectable (negative) to detectable (positive). The fitted survival after clinical diagnosis for cases that are ctDNA negative at blood draw is therefore a mixture of the ctDNA negative at diagnosis and ctDNA positive at diagnosis survival curves. We calculated the ratio of this mixture using the previous fitted model for sojourn time affecting detectability because in the CPS‐3 study, we have no direct observation of ctDNA positivity at clinical diagnosis. The output of this model is an estimate of the prognostic effect of ctDNA‐positive or ctDNA‐negative status at clinical diagnosis on the hazard ratio for survival, when compared to the expected survival of an individual of a given age and sex and a given cancer type and stage.

Cancer‐specific survival data for individuals in CPS‐3 were calculated as the first of either: (1) follow‐up after blood draw (up to 9 years) using cancer registry data or (2) time to mortality (up to 109 months) using death registry data. Data were followed up to a December 31, 2017, cutoff. SEER‐reference CSS data were taken from diagnosis years 2006 through 2015 in 17 SEER geographic areas, with follow‐up to November 2020, and stratified by age, sex, cancer type, and SEER Summary Stage. Individuals were matched to reference by cancer type without further subdivision into subtypes. SEER‐draw specifications are detailed in the Supplementary Information.

Fitting models

Models for detection and prognosis were fitted using the Stan Bayesian modeling language using the rstan package in R. ¹⁶ The probability of detection depends on study, cancer type and stage, and a time‐dependent rate also reliant on cancer type and stage.

For prognosis, this detection model was extended to fit a hazard ratio at clinical diagnosis that depends on both the observed ctDNA status at blood draw and the time‐dependent probability of changing ctDNA status between blood draw and diagnosis. In addition, a one‐parameter proportional hazards model was used to note the clinical significance of ctDNA positivity at blood draw. ¹⁷

We report posterior credible intervals and medians as approximate confidence intervals in the figures and tables.

RESULTS

Cancer signal detection rates decrease rapidly over time

We estimated preclinical sojourn times using the observed rate of cancer signal detection and how it varied based on the time between clinical diagnosis and blood draw. We further estimated any additional prognostic effect of ctDNA detection at clinical diagnosis on survival above and beyond known clinical factors. We report these effects here.

The average cancer detection rate stratified by yearly quarter since blood draw (CPS‐3) decreased with time. The fitted results followed the data trend (Figure 3). Comparing estimates derived from CCGA3 data (zero lead time) for the same cancer types and stages at each time after blood draw, we noted that fitted and observed detection rates resembled CCGA3 estimates at very short lead times (Figure 3). The fitted model included a term allowing for any deviation in the ability to detect ctDNA between studies. Translating the logit‐scale value of this term to the average detection rate for CCGA3 (50.9% in this subset) yielded a 95% credible interval [CI] for the difference of [–7.9%, 4.3%], which spanned zero.

FIGURE 3.

Observed detection rates (blue squares) within each 3‐month period change with time between blood draw and clinical diagnosis (with 95% pointwise credible intervals using Jeffrey’s method). Fitted line (green line connecting triangles) shows median estimate of mean signal detection rate from the fitted model for each period. Gray ribbon shows 95% posterior credible interval for the mean signal detection rate in each period. The expected detection rate at clinical diagnosis using sensitivities from CCGA3 (all at zero months of lead time) applied to the CPS‐3 cancers in each time window is shown as a visual reference (red line connecting circles). In each quarter, the expected detection rate at clinical diagnosis from CCGA3 fluctuates given the cancer composition, illustrating an additional source of expected variation. The additional difference with the observed detection rate between the red line (expected at clinical diagnosis) and the blue dots (observed at blood draw) can be explained by cancer progressing through its natural history and consequently changing detectability (ctDNA status) between blood draw and diagnosis. CCGA3 indicates Circulating Cancer Genome Atlas third substudy; ctDNA, circulating tumor DNA.

We verified that these trends in detection were consistent with natural history models. The observed detection rates closely resembled detection rates predicted by a previously published model, using the dwell time scenario referred to as “fast‐aggressive” in that publication (Figure 4; Supplementary Figure 1). ¹⁵

FIGURE 4.

In the model fitted to the observed data, the resulting estimates for the duration of the detectable window are short. However, the observed data match predicted detection rates and the change through time given the natural history assumptions for a fast‐aggressive tumor growth scenario of cancer of a previously published model. ⁷ This suggests that the observed data are largely consistent with the previously published model’s natural history assumptions and therefore is consistent with the modeled potential for reduction in late‐stage diagnosis using a ctDNA‐based multicancer early detection test. ctDNA indicates circulating tumor DNA.

Sojourn time varies widely by cancer type

The per‐cancer fitted timescale for sojourn time varied noticeably by cancer type. Median [CI] sojourn time by type ranged from 0.49 [0.26, 0.88] years for pancreas to 2.45 [1.14, 4.87] years for lymphoma (Figure 5). Posterior estimates have wide CIs when few informative observations were available for a cancer type. As stage increased, the estimated median per‐stage sojourn time increased (Supplementary Table 2); however, CIs were broad.

FIGURE 5.

Estimated sojourn times stratified by cancer (sorted by median). For example, pancreatic cancer has a shorter expected duration of shedding detectable ctDNA than lymphoma does. Intervals represent posterior 95% credible intervals around the median values. Number of detections and number of observations in CPS‐3 are noted by cancer type (in parentheses). Intuitively, cancer types that have high sensitivity observed at clinical diagnosis (i.e., in CCGA3) but are not detected at later time points in CPS‐3 drive estimates toward shorter sojourn times, cancer types that are detected at later time points drive estimates toward longer sojourn times. Cancer types with low expected sensitivity or few observations in CPS‐3 do not provide strong evidence toward any sojourn time and have wide intervals associated with them. Cancer types with fewer than five individuals in CPS‐3 have exact numbers censored. Because the time interval over which observations for cancer incidence after blood draw are taken is 3 years, these numbers should not be taken as reflecting episode sensitivity over any shorter time span. CCGA3 indicates Circulating Cancer Genome Atlas third substudy; CPS‐3, Cancer Prevention Study 3; ctDNA, circulating tumor DNA.

ctDNA status at blood draw indicates clinically significant cancers, with prognostic effect beyond clinical factors (age, sex, cancer type and stage)

ctDNA status at blood draw was associated with large differences in CSS. Stratified by cumulative year after blood draw (Supplementary Table 3), as all individuals positive at blood draw in an interventional experiment would be followed for some interval, the univariate hazard ratio for death based on positivity at blood draw was 7 in the first year but decreased to 3.5 when compared to populations in all 3 years of the study. These values contain large effects from differences in cancer type spectrum between populations (Supplementary Figure 2), as can be seen in the expected survival estimates from SEER (Figure 6). Kaplan–Meier computed survival estimates of CSS at 9 years from blood draw were 54% for ctDNA‐positive participants (slightly higher than the expected 46%) and 79.5% for ctDNA‐negative participants (comparable to the expected 75%) (Figure 6). Compared to population‐based reference survival rates for age, sex, cancer type and stage, and lead time in SEER data, CSS was better in each group, suggestive of healthy volunteer bias (see Methods and Figure 6).

FIGURE 6.

Survival data separated by ctDNA detection at blood draw. Cancer types, stages, and lead times are different in ctDNA‐positive and ctDNA‐negative cases; therefore, direct comparison is not recommended for prognostic purposes. In each case, a reference (solid) curve of expected survival via SEER data is computed given the cancer type, stage, and lead time. Because individuals have 100% probability of surviving to clinical diagnosis, the reference survival for each individual is 100% during lead time, after which reference survival from SEER is applied for the cancer type and stage. Because the probability of converting from ctDNA negative at blood draw to ctDNA positive at clinical diagnosis changes with lead time, the prognostic effect of ctDNA status at blood draw changes depending on the lead time. The fitted (dashed) curve adjusts the reference curve for this variable hazard ratio. The number at risk is included at the bottom of each graph. ctDNA indicates circulating tumor DNA; SEER, Surveillance, Epidemiology, and End Results.

Accounting for potential change in ctDNA detection between blood draw and clinical diagnosis, the estimated hazard ratio [CI] for ctDNA‐positive cancers at time of clinical diagnosis (as opposed to time of blood draw) was calculated to be 0.72 [0.61, 0.84] compared to reference survival data in SEER (Supplementary Figure 3). Furthermore, the fitted hazard ratio [CI] for ctDNA‐negative cancers at clinical diagnosis was estimated to be 0.36 [0.18, 0.62] (Supplementary Figure 3). Taking the ratio, we obtained a hazard ratio [CI] for positive versus negative as 1.98 [1.08, 4.22].

DISCUSSION

These results provide a glimpse into preclinical cancer biology of ctDNA shedding at detectable levels that inform the implementation of ctDNA‐based MCED screening programs at the population level. Here, we estimated the duration of the average preclinical detection window (sojourn time) in which ctDNA detection occurs across multiple cancer types and stages, including numerous cancer types with no current screening paradigm. These results and the natural history of the ctDNA signal are comparable to those predicted from a previously published model of cancer natural history and detection (Figure 4). ¹⁵ These findings are consistent with, although they do not confirm, models that imply that an annual, sustained MCED screening program may lead to a reduction in late‐stage cancer and a corresponding reduction in mortality. ¹⁵

These preclinical duration estimates have important implications for the outcomes of MCED screening trials. The first time a population is screened (e.g., the prevalent screen), there is a backlog of cancers to be detected that is proportional to the average preclinical duration of the detection window for each cancer type. Cancers with long detection windows, as observed here for lymphoma, may be overrepresented in the first screening rounds compared to faster‐growing solid tumors and may have reduced relative representation in later rounds of screening. Previous literature suggests that screening intervals should be similar to the average detection window ¹⁸ ; in this case, our analysis implies that an appropriate screening interval is approximately 1 year (Supplementary Table 2), as was previously suggested and modeled. ¹⁵

Taking lead time into account, preclinical ctDNA positivity was observed in individuals who went on to develop clinically significant cancers, limiting the potential that ctDNA‐detected cancers represent overdiagnosis. Despite developing clinically significant cancers, participants with ctDNA‐positive cancers at blood draw still had survival close to SEER for their cancer type and stage, even up to 9 years of follow‐up, consistent with previous reports with shorter follow‐up durations. ³ , ¹⁹ This includes early stages (Supplementary Figures 4 and 5), although data are limited by the number of early‐stage cancers. These results suggest that ctDNA‐positive cancers have outcomes similar to those expected for a comparable population‐based series of patients with cancer by type and stage and thus may be similarly amenable to treatment (e.g., do not look to have significantly worse than average outcomes). ¹ , ³

This study has several strengths. Importantly, the case‐cohort study design nested in a large prospective cancer cohort study with uniform follow‐up allowed us to study preclinical detection windows in an unbiased way impossible in case‐control study designs. Additionally, the MCED test used in this study used a locked classifier not trained on data from the studies, eliminating concerns of bias in feature selection or classifier training. The study population included all cancer types that arose in defined populations, allowing explicit modeling of sojourn times across multiple cancer types and stages. Joint modeling provides a linkage between results seen in a case‐control study and the results from a preclinical prospective cohort study. This in turn enables the separation of test sensitivity and natural history effects when examining observed detection rates as a function of time. Finally, we have up to 9 years of follow‐up to examine survival.

This study has certain limitations. We cannot use these data to estimate the length of time during which cancers are not detectable by ctDNA‐based methods because we cannot tell whether individuals have no detectable ctDNA or how long cancers persist without detectable ctDNA. In addition, because CPS‐3 was not an interventional study, the cancer status (including stage) at blood draw was not known; thus, potential reductions in late‐stage cancer diagnosis or cancer‐specific mortality cannot be examined in this dataset. Furthermore, as no blood was collected at time of cancer diagnosis in CPS‐3, direct observation of ctDNA status at clinical diagnosis is unavailable and must be modeled. Because clinical staging was different between the two studies, precision of results may be limited. In addition, individual cancer types may contain subtypes with differing properties (e.g., squamous vs adenocarcinoma lung; hormone receptor status in breast) not individually modeled. The CPS‐3 population also consisted of younger individuals, many of whom were younger than typical cancer screening ages and the intended‐use population of the MCED test. As observed in the CPS‐3 cohort study, ¹⁰ , ¹³ there is a high probability of healthy volunteer bias affecting the cancer incidence and potentially survival, leading to lower cancer incidence and mortality rates than the general (SEER) population. Unobserved biases in clinical cancer diagnosis, such as length time bias, overdiagnosis, and healthy volunteer bias, may also correlate with ctDNA status and be confounded with prognosis.

These results require confirmation in future prospective analyses. Although modeled differences in inherent detectability between studies were not significantly different from zero, there were differences in sample handling and storage between the two studies, which may reduce test performance by shortening timescales for cancer detection and/or decreasing sensitivity. Furthermore, timescales and prognoses measured in this younger‐than‐typical screening age population may differ from those in an older, screening‐age population. Finally, analyses with more data may examine more complex models for sojourn time and shedding of detectable levels of ctDNA.

These results inform design of future MCED screening studies and programs, consistent with an annual screening interval. Despite rapid cancer progression (Figure 5), we detected a noticeable fraction of positive cancers that would have been clinically diagnosed more than 1 year in the future. Censoring at 1‐year follow‐up would have caused these noninterventional positives to either never be sampled for testing, or worse, appear to be false positives. Differences in the size of the detectable window will affect the number of cancer types detected in a single blood draw in unscreened populations, thereby affecting estimates of episode sensitivity (the fraction of cancers found through screening in a given interval, or episode). Furthermore, the rapid natural history of cancers affects estimates of episode sensitivity when compared to test sensitivity. ²⁰ All of these factors must be considered when designing, modeling, and interpreting future studies.

CONCLUSIONS

In summary, these cohorts provided a rich dataset enabling estimation of both preclinical sojourn time and prognosis. The rapid timescales inferred from these data suggest that annual screening may be useful for MCED screening, as previously suggested. ¹⁵ Consistent with previous studies, survival was better than expected for SEER in both subgroups when estimated at 9 years of follow‐up. This improvement in survival held even when restricting the model to cases that were ctDNA positive, indicating these cases were as or more survivable than expected for their cancer type and stage. These results suggest that cancer screening through ctDNA‐based assays may allow for survival benefits to early detection.

AUTHOR CONTRIBUTIONS

Earl Hubbell: Conceptualization; methodology; software; supervision; validation; visualization; writing—original draft; and writing—review & editing. Alpa V. Patel: Conceptualization; funding acquisition; methodology; resources; supervision; and writing—review & editing. Christina A. Clarke: Conceptualization; funding acquisition; methodology; resources; writing—original draft; and writing—review & editing. Emily Deubler: Methodology; resources; software; and writing‐review & editing. Eric T. Fung: Conceptualization; project administration; resources; supervision; and writing—review & editing. Rong Jiang: Data curation; software; and writing—review & editing. Allison W. Kurian: Writing—review & editing. Cari Lichtman: Methodology; resources; software; writing—original draft; and writing—review & editing. Lauren R. Teras: Conceptualization; methodology; resources; and writing—review & editing. Oliver Venn: Conceptualization; methodology; and writing—review & editing. Nan Zhang: Conceptualization; supervision; and writing—review & editing. Charles Swanton: Conceptualization; supervision; and writing—review & editing.

CONFLICT OF INTEREST STATEMENT

Earl Hubbell reports employment at GRAIL, Inc.; stock and other ownership interests in Illumina and GRAIL, Inc; patents, royalties, and other intellectual property with GRAIL, Inc. Alpa V. Patel reports no conflicts of interest. Christina A. Clarke reports employment at GRAIL, Inc.; stock and other ownership interests in Illumina and GRAIL, Inc. Emily L. Deubler reports no conflicts of interest. Eric T. Fung reports employment at GRAIL, Inc.; stock and other ownership interests in GRAIL, Inc.; patents, royalties, and other intellectual property with GRAIL, Inc. Rong Jiang reports former employment at GRAIL, Inc.; stock and other ownership interests in Illumina and GRAIL Inc. Allison W. Kurian reports unfunded research collaborations with Ambry Genetics, Color Health, Bioreference/GeneDx, Labcorp, Myriad, Exact Sciences, Gilead, Genentech, Merck, Foundation, and Tempus. Cari Lichtman reports no conflicts of interest. Lauren R. Teras reports no conflicts of interest. Oliver Venn reports employment at GRAIL, Inc.; stock and other ownership interests in Moderna Therapeutics, Illumina, and GRAIL, Inc; patents, royalties, and other intellectual property with GRAIL, Inc. Nan Zhang reports former employment with GRAIL, Inc.; stock and other ownership interests in Illumina and GRAIL, Inc.; pending patent with GRAIL, Inc., on the microsimulation methodology to evaluate multicancer early detection tests. Charles Swanton reports stock and other ownership interests in Achilles Therapeutics, Apogen Biotechnologies, Bicycle Therapeutics, and Epic Sciences; honoraria from Amgen, AstraZeneca, Boehringer Ingelheim, Bristol Myers Squibb, Celgene, GlaxoSmithKline, Illumina, Lilly, MSD Oncology, Novartis, Ono Pharmaceutical, Pfizer, Roche, and Roche/Genentech; consulting or advisory role for Achilles Therapeutics, Amgen, AstraZeneca, Bicycle Therapeutics, Bristol Myers Squibb, Genentech/Roche, GlaxoSmithKline, GRAIL, Inc., Illumina, Medicxi, Metabomed, MSD, Novartis, Pfizer, Roche, and Sarah Cannon Research Institute; research funding from Archer, AstraZeneca, BMS, Boehringer Ingelheim, Ono Pharmaceutical, Personalis, Pfizer, and Roche; patents, royalties, other intellectual property: Founder of Achilles Therapeutics, a biotechnology company funded by Syncona/Wellcome Trust to target clonal neoantigens through vaccine and cell therapy approaches; immune checkpoint intervention in cancer (PCT/EP2016/071471); method for determining whether an HLA allele is lost in a tumor (PCT/GB2018/052004); method for identifying responders to cancer treatment (PCT/GB2018/051912); method for treating cancer (PCT/EP2016/059401); method for treating cancer based on identification of clonal neo‐antigens (PCT/EP2016/059401); method of detecting tumor recurrence (PCT/GB2017/053289); method of identifying insertion/deletion mutation targets (PCT/GB2018/051892); method of predicting survival rates for patients with cancer (PCT/GB2020/050221); methods for lung cancer detection (PCT/US2017/028013); and uncompensated relationships with AstraZeneca (institutional).

Supporting information

Supplementary Material

DATA AVAILABILITY STATEMENT

Code, along with a synthetic dataset with similar distribution to the original dataset to demonstrate analysis procedure, are available at https://github.com/grailbio‐publications/Hubbell_Joint_Model. The original dataset is available from the American Cancer Society by following the ACS Data Access Procedures (https://www.cancer.org/content/dam/cancer‐org/research/epidemiology/cancer‐prevention‐study‐data‐access‐policies.pdf) for researchers who meet the criteria for access to confidential data.