Real-World Validity of Tissue-Agnostic Circulating Tumor DNA Response Monitoring in Lung Cancers Treated With Chemotherapy, Immunotherapy, or Targeted Agents

Abstract

Introduction

Circulating tumor DNA (ctDNA) monitoring is emerging as a minimally invasive complement to tumor imaging. We evaluated the validity of tissue-agnostic ctDNA quantification across four treatment modalities in NSCLC and SCLC.

Methods

Data from consenting patients were collected from electronic health records as part of the Prospective Clinico-Genomic study (NCT04180176). ctDNA tumor fraction (TF) was retrospectively calculated for plasma collected six to 15 weeks after therapy initiation. TF dynamics were compared among an exploratory cohort, NSCLC and SCLC validity cohorts, and by therapy class.

Results

In on-treatment plasma, undetectable TF was associated with longer real-world progression-free survival and real-world overall survival in exploratory (21.8 versus 8.8 mo; hazard ratio [HR] = 0.35, 95% confidence interval [CI]: 0.24–0.50), validity NSCLC (23.5 versus 9.5 mo; HR = 0.34, 95% CI: 0.22–0.53), and validity SCLC (15.9 versus 8.3 mo; HR = 0.19, 95% CI: 0.08–0.42) cohorts. Equal to or greater than 90% and equal to or greater than 50% TF reduction from baseline was also associated with significantly improved outcomes. ctDNA dynamics differed by treatment class: TF reported greater discriminatory power for selecting tumor responses to immunotherapy and targeted therapy (≥50% decrease in 91% of responders versus 24% of nonresponders) than chemotherapy and chemo-immunotherapy (86% versus 60%). TF dynamics correlated with outcomes, but models of real-world progression-free survival and real-world overall survival were improved when tumor response was included.

Conclusions

Tissue-agnostic monitoring of molecular response on the basis of ctDNA TF dynamics has utility in the real-world setting across four different treatment regimens. These results suggest that ctDNA dynamics may be complementary to tumor imaging in both NSCLC and SCLC to better inform patient care.

Introduction

Genomic profiling of circulating tumor DNA (ctDNA) is an established testing modality that is routinely incorporated into standard of care for patients with advanced NSCLC and is emerging in SCLC.¹ Although tissue-based methods remain the accepted standard, liquid biopsies provide a practical alternative for patients whose tumor biopsies are unavailable or difficult to obtain² and serve as comprehensive assessments of a patient’s entire metastatic burden.³ Several Food and Drug Administration–approved liquid biopsy companion diagnostics are available today for physicians to aid treatment decisions.4, 5, 6, 7 Beyond therapy selection, the prognostic value of pretreatment ctDNA levels has been reported in lung, breast, colorectal, prostate, urothelial, and other cancers, which suggests that risk-stratifying on ctDNA detection can help tailor therapeutic interventions for a patient’s disease.8, 9, 10, 11, 12, 13

Monitoring of ctDNA is emerging as a minimally invasive complement to tumor imaging for assessing treatment effect. Several studies have retrospectively analyzed results from single institutions or randomized clinical trials and shown the utility of ctDNA monitoring in patients treated with immune checkpoint inhibitors (ICIs) or targeted therapies.14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24 There is some limited evidence to suggest that ctDNA monitoring may also have utility for patients with NSCLC treated with platinum-based chemotherapy.²⁵ In addition, data supporting ctDNA monitoring in patients with SCLC is promising, though the number of patients evaluated has been low.^26,27 ctDNA monitoring may be particularly useful for patients treated with ICI, whose radiographic assessments of response can be unclear, specifically in differentiating between long- and short-term stable disease or identifying pseudoprogression.^28,29

Although studies have been presented with the aim of identifying an optimal cutoff for defining ctDNA response, it remains unclear whether cutoffs differ by therapy or indication. Here, we analyzed plasma collected from patients with NSCLC or SCLC in the real-world Prospective Clinico-Genomic study (PCG; NCT04180176) to evaluate the validity of ctDNA monitoring in patients treated with diverse therapies in a real-world setting and to elucidate the differences in ctDNA dynamics across therapy modalities.

Materials and Methods

PCG Study Design

In short, PCG was a multicenter study that enrolled patients with a documented diagnosis of metastatic NSCLC or extensive-stage SCLC who were initiating treatment with any line of standard-of-care therapy.³⁰ This study leveraged prospectively collected, longitudinal electronic health record data, comprising de-identified patient-level structured and unstructured data. Patients were sourced from 23 participating sites from the Flatiron Health Research Network, which included approximately 280 cancer clinics in the United States (∼800 sites of care). The PCG data model leveraged many of the data fields incorporated in the nationwide de-identified Flatiron Health database and Flatiron Health-Foundation Medicine clinico-genomic database.31, 32, 33

Patients were eligible to have plasma collected at three time points: (1) trial enrollment (0–28 days before initiating a new line of therapy), (2) during treatment (if no progression at the first tumor assessment or with progression at the first tumor assessment but did not end treatment within 14 days after that assessment), and (3) progression or end-of-therapy if they experienced progression or ended treatment, or if the progression was observed at the first tumor assessment and treatment was discontinued within 14 days (Fig. 1A).³⁰

Figure 1.

Study description and patient inclusion. (A) The PCG study paired electronic health record data from consenting patients with comprehensive genomic profiling results from blood and tissue, as available.³⁰(B) Patients from the PCG trial were included in three overlapping cohorts on the basis of prespecified criteria. Patients tested with the prototype version of the LBx assay or patients with NSCLC treated with non-ICI regimens were analyzed in an exploratory cohort. (C) For sub-analysis within each cohort, the proportion of patients included in Cohort A but not Cohort B was similar for the Exploratory and NSCLC validity cohorts. Nevertheless, a smaller proportion of patients within the SCLC validity cohort were included in Cohort A but not Cohort B. (D) Cohorts were overlapping such that all patients in Cohort B were included in Cohort A and Cohort C. An additional 75 patients who were part of Cohort A but not B were included along with 117 additional patients who were excluded from the outcomes analyses. #, number; ctDNA, circulating tumor DNA; EHR, electronic health records; FMI, Foundation Medicine; ICI, immune checkpoint inhibitor; LBx, liquid biopsy; PCG, Prospective Clinicogenomic; pts, patients; rwORR, real-world overall response rate; rwOS, real-world overall survival; rwPFS, real-world progression-free survival; SOC, standard of care.

This study was conducted in conformance with the International Council for Harmonization E6 guideline for Good Clinical Practice, the principles of the Declaration of Helsinki, and the Code of Federal Regulations on the Protection of Human Subjects. All patients provided informed consent.

ctDNA Analysis and ctDNA TF

ctDNA analysis was conducted using FoundationOne Monitor, a tissue-agnostic ctDNA monitoring assay using hybrid capture next-generation sequencing, leveraging the same sequencing platform as FoundationOne Liquid CDx assay.³⁴ ctDNA TF was calculated through previously described methods.^35,36 In brief, ctDNA TF was quantified by conducting a comprehensive evaluation of genomic alterations, including aneuploidy, short-variants, and certain rearrangements, while excluding mutations and aneuploidy derived from clonal hematopoiesis (CH) using patterns related to the sizes of cell-free DNA fragments (Supplementary Fig. 1).

Patients with plasma collection from September 2019 to October 2020 (48% of patients) had ctDNA testing performed on a prototype ctDNA monitoring assay (70-gene panel). The analysis of these samples was exploratory and was not accounted for during analysis planning. Patients in the validity cohorts had plasma collected from November 2020 to August 2023 and had ctDNA testing performed on FoundationOneMonitor.

Somatic and Germline Predictions in Tissue Biopsies

For a subset of patients, tissue genomic profiling was performed using hybrid capture next-generation sequencing (Food and Drug Administration–approved FoundationOne CDx), as previously described.³⁷ For detected variants, a patient-specific somatic or germline origin was computationally inferred using a previously described algorithm.³⁸ In validation testing against matched normal specimens, accuracy was 95% and 99% for somatic and germline variants, respectively.

Cohort Selection

Patients from the PCG study were selected for evaluation here on the basis of several criteria ensuring data availability and quality (Fig. 1B). Two primary, prespecified analyses were performed: first, real-world overall survival (rwOS), real-world progression-free survival (rwPFS), and real-world overall response rate (rwORR) were assessed, segmented by on-treatment ctDNA detection irrespective of ctDNA levels at enrollment (cohort A; Supplementary Fig. 2A). Second, rwOS, rwPFS, and rwORR were compared by levels of ctDNA change from enrollment to on treatment (cohort B; Supplementary Fig. 2B). Prespecified cutoffs of 90%³⁹ and 50%⁴⁰ decrease were used to define molecular response (MR). Patients with ctDNA not detected (ND) at either time point were withheld and analyzed separately. Cohorts A and B included patients with liquid biopsies collected six to 15 weeks after the start of therapy, regardless of the timing of their first tumor assessment or whether the specimen was collected at disease progression. Cohort B included all patients in cohort A with enrollment specimens collected no more than two weeks before therapy start and a quantifiable ctDNA change. A similar proportion of patients were included in cohort A but not cohort B in the exploratory (21%) and NSCLC validity (28%) cohorts, whereas a lower percentage of patients in the SCLC validity cohort were included in cohort A only (13%) (Fig. 1C). Of the 27 patients who were exclusively included in cohort A (Fig. 1D), 19 were from the exploratory cohort, five were from the NSCLC validity cohort, and three were from the SCLC validity cohort (data not shown).

For both primary analyses, cohorts A and B were each further divided into three subgroups: (1) The exploratory cohort included all patients tested with the prototype assay and any patients with NSCLC treated with non-ICI regimens; (2) the NSCLC validity cohort included patients with NSCLC treated with ICI-containing regimens and tested with the analytically validated assay; and (3) the SCLC validity cohort included patients receiving all regimens and tested with FoundationOneMonitor (Fig. 1B).

All patients with enrollment liquid biopsy before the start of therapy (i.e., more permissive than cohorts A and B) and an on-treatment liquid biopsy collected at least three weeks after treatment initiation (also more permissive) were chosen for exploratory genomics analysis as cohort C (Supplementary Fig. 2C). All patients from cohort B, 75 patients who were in cohort A but not cohort B, and an additional 117 patients (total 540 patients) were included (Fig. 1D). Of the 75 patients who were exclusively included in cohort A and cohort C but not cohort B, 32 were from the exploratory cohort, 39 were from the NSCLC validity cohort, and four were from the SCLC validity cohort (data not shown).

Real-World Outcomes

rwPFS was calculated from the initiation of treatment until either an abstracted real-world progression event or death, whichever came first. Patients without progression or death at their last clinic visit were right-censored. rwOS was calculated from treatment initiation to death from any cause, and patients with no record of death were censored at the last electronic health record structured activity date. Risk set adjustment was applied to account for left truncation. Patients entered the at-risk population at the time of their second liquid biopsy specimen collection. Truncation independence with censoring was evaluated with Kendall’s tau, with a p value of 0.05 or less considered acceptable.

Hazard ratios (HRs) were calculated using multivariate Cox proportional hazard models adjusting for relevant covariates (e.g., sex, age, race, smoking history, Eastern Cooperative Oncology Group performance status, liver and central nervous system involvement, histologic subtype, and programmed death-ligand 1 status). Variables with less than 25% missingness were handled by simple imputation with expected values determined using random forests with the R package “missForest.” In Cox proportional hazard models, imputed values were treated identically to measured values. For variables with more than 25% missingness, a missing value was included. Variables with fewer than five observations were excluded.

Real-world response to therapy (rwR) was abstracted from unstructured clinician documentation as described and validated previously.⁴¹ The rwORR was calculated as the percentage of patients who achieved complete response or partial response.

Statistical Analysis

Differences in rwPFS and rwOS were assessed with the log-rank test and HRs from Cox proportional hazard models and visualized with Kaplan-Meier analysis. Fisher’s tests and Wilcoxon rank-sum tests were used to assess differences between cohorts of categorical and continuous variables, respectively. Harrell’s concordance index (C-index) is a measure of model discriminatory performance, with values from 0.5 to 1.0, in which 0.5 reflects a completely random result and 1.0 reflects perfect prediction. It was used as generated from the “survival” package in R (RRID:SCR_021137). Statistical significance was assessed from 95% confidence intervals (CIs). Linear relationships between continuous variables were assessed with Pearson’s correlation coefficient.

Statistics, computation, and plotting were carried out using R 4.3.2 (Posit, Boston, MA, RRID:SCR_001905) packages ggplot2 (RRID:SCR_014601), survminer (RRID:SCR_021094), survival (RRID:SCR_021137), missForest (RRID:SCR_018543), TransSurv, and tidyverse (RRID:SCR_019186).

Results

On-Treatment Detection of ctDNA

In total, 450 patients (cohort A) were assessed for on-treatment ctDNA status. Patient characteristics for the three primary analysis subgroups can be seen in Supplementary Table 1. Within the exploratory subgroup, on-treatment ctDNA detection was associated with shorter median rwPFS (ctDNA TF ND: 8.5 mo versus detected [D]: 3.9 mo; HR = 0.42, 95% CI: 0.30–0.58, log-rank p < 0.0001; Fig. 2A, Supplementary Fig. 3A) and shorter rwOS (ND: 21.8 mo versus D: 8.8 mo; HR = 0.35, 95% CI: 0.24–0.50, p < 0.0001, Fig. 2B, Supplementary Fig. 3B). For patients in the NSCLC validity cohort with on-treatment ctDNA detection, we found a significant difference in median rwPFS of 3.3 months versus 9.8 months for patients with ND (HR = 0.26, 95% CI: 0.18–0.40, p < 0.0001; Fig. 2C, Supplementary Fig. 3C) and rwOS (ND: 23.5 mo versus D: 9.5 mo; HR = 0.34, 95% CI: 0.22–0.53, p < 0.0001; Fig. 2D, Supplementary Fig. 3D). In the SCLC validity cohort, we also found a significantly longer median rwPFS among patients with ctDNA TF ND (6.1 mo) compared with those with TF D (4.3 mo) (HR = 0.39, 95% CI: 0.19–0.80, p = 0.01; Fig. 2E, Supplementary Fig. 3E). In this cohort, rwOS was superior as well (ND: 15.9 mo versus D: 8.3 mo; HR = 0.19, 95% CI: 0.08–0.42, p < 0.0001; Fig. 2F, Supplementary Fig. 3F).

Figure 2.

Patient outcomes assessed by ctDNA detection status 6 to 15 weeks after the start of therapy. In an exploratory analysis, ctDNA detection was associated with inferior (A) rwPFS and (B) rwOS. These findings were repeated in prespecified validation cohorts consisting of (C and D) patients with NSCLC treated with ICI and (E and F) patients with SCLC. Gray box in rwOS curves marks the time period when no patients are at risk (0–6 wk from the start of therapy). CI, confidence interval; ctDNA, circulating tumor DNA; ICI, immune checkpoint inhibitor; OS, overall survival; PFS, progression-free survival; rwOS, real-world overall survival; rwPFS, real-world progression-free survival.

In the NSCLC validity cohort, the difference in rwORR between patients with ctDNA D and those without was significant (ND: 77% versus D: 29%; OR = 0.13, 95% CI: 0.05–0.28). Nevertheless, in the SCLC validity cohort, no significant difference in rwORR was observed (ND: 84% versus D: 75%; OR = 0.58, 95% CI: 0.11–2.68; Supplementary Table 2). In addition, we evaluated the response among patients above or below a ctDNA TF cutoff of 1%. Five patients in the SCLC validity cohort had ctDNA TF D but below 1%. Interestingly, all five of these patients with low TF achieved a partial response. In total, 20 such patients were observed in the NSCLC validity cohort, but only nine of these patients had a complete or partial response (Supplementary Table 2). Comparisons in patient characteristics between patients with ctDNA D and ND are shown in Supplementary Table 3.

Given the real-world nature of the PCG study, the timing of on-treatment plasma collection was variable. We selected a range of 6 to 15 weeks after therapy start to mirror prior studies that have analyzed plasma collected at any point between cycle 3 day 1 and cycle 5 day 1.⁴² In this data set, outcomes were similar for patients with specimens collected within the inclusion window. Furthermore, specimens were frequently collected within approximately two weeks of first tumor assessment as specified in the trial design (Supplementary Table 4).

ctDNA Dynamics From Enrollment to On-Treatment Liquid Biopsy

Beyond on-treatment ctDNA detection, we sought to understand the relationship between patient outcomes and ctDNA change from pretreatment baseline. Characteristics for 245 patients included in these analyses (cohort B) are shown in Supplementary Table 5. Using a prespecified cutoff of 90% decrease to define MR, longer median rwPFS (MR: 7.7 mo versus no MR: 3.1 mo; HR = 0.24, 95% CI: 0.16–0.36; Fig. 3A, Supplementary Fig. 4A) and rwOS (MR: 17.6 mo versus no MR: 8.3 mo, HR = 0.29, 95% CI: 0.19–0.44; Fig. 3B, Supplementary Fig. 4B) was observed in the exploratory subgroup. As seen with on-treatment ctDNA detection, these findings held in the NSCLC validity cohort, with a median rwPFS among patients with MR of 8.6 months versus 2.8 months among those with no MR (HR = 0.15, 95% CI: 0.06–0.37; Fig. 3C, Supplementary Fig. 4C). The median rwOS in the NSCLC validity cohort was also significantly longer in patients with MR (19.4 mo) versus no MR (8.4 mo) (HR = 0.18, 95% CI: 0.07–0.50; Fig. 3D, Supplementary Fig. 4D). The SCLC validity cohort also had significant differences in median rwPFS (MR: 5.6 mo versus no MR: 4.5 mo; HR = 0.33, 95% CI: 0.14–0.80; Fig. 3E, Supplementary Fig. 4E) and rwOS (MR: 15.7 mo versus no MR: 7.2 mo; HR = 0.11, 95% CI: 0.04–0.33; Fig. 3F, Supplementary Fig. 4F). Comparisons in patient characteristics between patients with MR and those without are shown in Supplementary Table 6.

Figure 3.

Patient outcomes assessed by molecular response (≥90% decrease in ctDNA from enrollment). (A) rwPFS and (B) rwOS in the exploratory cohort. (C) rwPFS and (D) rwOS in patients with NSCLC treated with ICI with or without chemotherapy. (E) rwPFS and (F) rwOS patients with SCLC. Gray box in rwOS curves marks the time period when no patients are at risk (0–6 wk from the start of therapy). CI, confidence interval; ctDNA, circulating tumor DNA; ICI, immune checkpoint inhibitor; OS, overall survival; PFS, progression-free survival; rwOS, real-world overall survival; rwPFS, real-world progression-free survival.

MR was also associated with higher rwORR in all three subgroups, though there were several patients with less dramatic decreases in ctDNA who responded to therapy (Fig. 4, Supplementary Tables 7 and 8). To assess the impact of smaller decreases in ctDNA, we evaluated a 50% decrease as a potential cutoff for MR in both validity cohorts. The association between MR and rwPFS and rwOS did not differ significantly with a cutoff of 50% as reported by overlapping C-index CIs (Supplementary Figs. 5 and 6). In both NSCLC and SCLC subgroups, reducing the MR cutoff to a decrease of 50% captured more patients with partial response, but since additional patients with stable or progressive disease were also classified as molecular responders, no improvement in rwORR was seen (Supplementary Table 7).

Figure 4.

Molecular response across exploratory and validity cohorts, as compared with best radiographic rwR. Molecular response after 6 to 15 weeks (≥90% decrease) was associated with superior best rwR in all three cohorts. Several patients with less pronounced decreases in ctDNA responded to therapy. ctDNA, circulating tumor DNA; rwR, real-world response.

We next assessed the outcomes of patients without ctDNA detection at both enrollment and on treatment (persistent negative) and compared them to patients without MR. As expected, baseline levels of ctDNA were higher in patients with SCLC than those with NSCLC (Supplementary Fig. 7). SCLC was excluded from this analysis as only two patients with SCLC had persistent negative ctDNA. Although rwORR response rates for patients with NSCLC with persistent negative ctDNA (64%) were lower than in patients with MR (83% for ≥90% decrease and 76% for ≥50% decrease; Supplementary Table 7), risk of progression and death for these patients relative to patients without MR was similar (Supplementary Fig. 8A). For patients treated with chemotherapy plus ICI, the median rwPFS and rwOS were both numerically but not significantly longer for patients with persistent negative ctDNA versus equal to or greater than 90% clearance in the exploratory cohort. Kaplan-Meier analyses found substantial overlap, and the risk of hazards was not significantly different (Supplementary Fig. 8B).

Combining MR With Real-World Imaging

To understand if MR is additive to traditional imaging, we grouped patients by radiographic response and ctDNA response. In all patients with NSCLC (exploratory and validity), those with both complete or partial response and MR had substantially better outcomes than other patients (rwPFS = 10.4 mo; rwOS = 28.2 mo; Fig. 5A), whereas patients with neither radiographic nor MR had the worst outcomes (rwPFS = 2.7 mo; rwOS = 8.2 mo; Fig. 5A). MR was associated with superior rwPFS and rwOS among patients with a documented rwR (rwPFS HR = 0.31, 95% CI: 0.17–0.55; rwOS HR = 0.20, 95% CI: 0.10–0.41) and those with no rwR (rwPFS HR = 0.22, 95% CI: 0.11–0.45; rwOS HR = 0.47, 95% CI: 0.24–90). When limiting to patients with NSCLC treated with chemotherapy plus ICI, the differences in patients whose radiographic and MRs were discordant trended toward favoring patients with MR but no rwR (median rwPFS–MR: 6.6 versus 3.8 mo; median rwOS: 12.3 versus 8.2 mo; Fig. 5B). Again, we assessed the association of MR and outcomes; within both responders (rwPFS HR = 0.28, 95% CI: 0.11–0.71; rwOS HR = 0.10, 95% CI: 0.03–0.31) and nonresponders (rwPFS HR = 0.05, 95% CI: 0.01–0.28; rwOS HR = 0.04, 95% CI: 0.01–0.27) results were consistent with the whole NSCLC cohort. In contrast, for patients with SCLC, MR was not associated with improved outcomes in patients with no documented rwR (rwPFS HR = 0.55, 95% CI: 0.17–1.73; rwOS HR = 0.73, 95% CI: 0.22–2.42), although the cohort was small. Among patients with a rwR, those who achieved MR had superior rwPFS (HR = 0.39, 95% CI: 0.20–0.73; Fig. 5C). The rwOS benefit (HR = 0.18, 95% CI: 0.09-0.39; Fig. 5C) was particularly pronounced in patients with rwR and MR compared with patients with rwR but no MR, who had median rwOS survival comparable to patients without a documented rwR (rwR & no MR: 8.3 mo; no rwR & MR: 8.7 mo; no rwR & no MR: 8.4 mo; Fig. 5C).

Figure 5.

MR and rwR. (A) In NSCLC, the combination of MR (≥90% decrease in ctDNA) and real-world response identified patients with the greatest benefit from therapy. (B) When limited to patients treated with ICI plus chemotherapy, MR trended toward a stronger correlation with rwPFS and rwOS, suggesting that ctDNA could be used in clinical trials in which novel agents are compared with standard of care ICI plus chemotherapy. (C) In patients with SCLC, a marginal rwPFS benefit was seen for patients with MR but no real-world response. Nevertheless, a large overall survival benefit was seen for patients with both molecular and rwRs. Gray box in rwOS curves marks time period when no patients are at risk (0–6 wk from start of therapy). CI, confidence interval; ctDNA, circulating tumor DNA; ICI, immune checkpoint inhibitor; ICI, immune checkpoint inhibitor; MR, molecular response; NR, not reported; OS, overall survival; rwOS, real-world overall survival; rwPFS, real-world progression-free survival; rwR, real-world response.

A similar analysis of patients with NSCLC treated with ICI monotherapy was confounded by small numbers of patients in the two discordant categories (Supplementary Fig. 9A) and for the subset of patients with NSCLC receiving chemotherapy as monotherapy concordant MR and imaging results were indicative of particularly good or poor prognosis, and discordant results fell between (Supplementary Fig. 9B).

Finally, we sought to understand the interaction between MR and rwR with respect to rwOS. In all patients with NSCLC, MR remained independently associated with improved rwOS (HR = 0.50, 95% CI: 0.28–0.89; Supplementary Fig. 10A), whereas no association between the lack of rwR and improved survival was observed (HR = 0.81, 95% CI: 0.44–0.1.51; Supplementary Fig. 10A). Similar results were seen in patients with NSCLC treated with chemotherapy plus ICI (HR = 0.26, 95% CI: 0.10–0.69; Supplementary Fig. 10B), whereas in patients with SCLC, MR was not associated with improved rwOS when accounting for the interaction with rwR (HR = 0.64, 95% CI: 0.21–1.96; Supplementary Fig. 10C). The interaction between MR and rwR was not significant in either NSCLC group (All NSCLC: HR = 0.51, 95% CI: 0.22–1.17; chemotherapy + ICI NSCLC: HR = 0.45, 95% CI: 0.12–1.66; Supplementary Fig. 10) but for patients with SCLC (HR = 0.29, 95% CI: 0.08–1.02; Supplementary Fig. 10), the trend mirrored the observations in Figure 5 and suggests that MR may have stronger discriminatory power for predicting survival in patients with rwR.

ctDNA Dynamics by Therapy Class

Next, we explored how levels of ctDNA change for real-world responders (complete or partial response) and nonresponders (stable or progressive disease) within both NSCLC and SCLC. High levels of ctDNA decrease were more common for those with real-world response in both NSCLC and SCLC. The overall distributions were similar, but increases in ctDNA were rare in SCLC, possibly due to the high levels of baseline ctDNA seen in these patients (Fig. 6A).

Figure 6.

Changes in ctDNA by therapy class. (A) Depth of ctDNA response was greater in patients who had a documented tumor response in both patients with SCLC and NSCLC. (B) Within NSCLC, levels of ctDNA change differed by therapy. In patients without a documented tumor response, decreases of 50% were rare in patients treated with ICI monotherapy or targeted therapy, whereas a 50% ctDNA decrease was common among nonresponders treated with chemotherapy. (C) A ctDNA decrease of 90% was associated with tumor response in all four treatment modalities. Patients whose ctDNA was persistently negative reported no association with response. (D) Regardless of treatment, increases in ctDNA were strongly associated with inferior rwPFS. Immunotherapy as monotherapy and targeted therapy were combined owing to similar response characteristics and relatively low patient numbers in each category. Chemo, chemotherapy; CI, confidence interval; ctDNA, circulating tumor DNA; IO, immuno-oncology; NS, not significant; rwPFS, real-world progression-free survival; TF, tumor fraction.

Considering findings from Friends of Cancer Research and collaborators on ctDNA responses to ICIs and tyrosine kinase inhibitors, we hypothesized that different therapy classes would exhibit different reductions in ctDNA.^40,43 We evaluated ctDNA changes within specific therapy classes for patients with NSCLC. For the 53 patients lacking a tumor response who were treated with chemotherapy alone or in combination with ICI, most (n = 32), nevertheless, reported a ctDNA decrease of at least 50%. For the 17 patients treated with either ICI monotherapy or targeted therapy without documented tumor response, only four patients saw a ctDNA decrease of at least 50% (Fig. 6B). A ctDNA decrease of any magnitude was only observed in one patient treated with targeted therapy who did not have a real-world response, though numbers were low (n = 9; Fig. 6C). Regardless of therapy, increases in ctDNA were associated with a significantly higher risk of real-world progression (Fig. 6D).

Variant Allele Frequency and ctDNA TF

We sought to understand the relationship between variant allele frequency (VAF) and ctDNA TF between longitudinal samples from patients in cohort C. Because ctDNA TF can be derived from either aneuploidy (copy number [CN] changes) or VAF after fragmentomic filtering, we separately analyzed cases in which ctDNA TF at both estimates were VAF based (VAF-VAF, n = 153) from those in which aneuploidy was used at enrollment or on treatment (n = 185). When comparing percent change in VAF to percent change in TF for alterations in common oncogenes (EGFR, KRAS, PIK3CA) a strong linear association was observed (Pearson’s correlation coefficient (r); All r = 0.864; VAF-VAF r = 0.914; CN r = 0.800; Supplementary Fig. 11A). The same comparison was made for genes often associated with CH (DNMT3A, ASXL1, TET2), and change in VAF reported no correlation to change in ctDNA TF (All r = -0.134; VAF-VAF r = -0.013; CN r = -0.271); Supplementary Fig. 11B). In tumor suppressor genes often seen in lung cancer, results are inconsistent. Changes in TP53 VAF, which can be CH-derived and tumor-derived,⁴⁴ had a modest correlation with ctDNA TF changes (All r = 0.353; VAF-VAF r = 0.322; CN r = 0.390; Supplementary Fig. 11C), consistent with the mixed origin of these alterations. RB1 and STK11 reported a correlation similar to common lung oncogenes (All r = 0.894; VAF-VAF r = 0.934; CN r = 0.866; Supplementary Fig. 11C), suggesting alterations in these genes rarely originate from CH.

Tissue-based profiling was also performed at enrollment with 130 patients. To assess the performance of ctDNA TF compared with an ad-hoc tissue-informed approach developed for this study, changes in VAF from variants D in tissue were compared with changes in ctDNA TF. As expected, percent change in VAF for variants which were computationally predicted to be somatic had a stronger association with changes in ctDNA TF than variants that were predicted to be germline (ρ = 0.857 versus 0.423; Supplementary Fig. 12).

Discussion

Deciding when to cease, intensify, or de-intensify treatment for lung cancer carries significant importance in prolonging patient outcomes and maximizing patient quality of life. Imaging remains the gold standard method in assessing treatment response, but ctDNA monitoring has emerged as a potential tool to enhance a physician’s decision making in the rapidly evolving landscape of immune and combination therapies.

Here, we report the real-world clinical validity of tissue-agnostic ctDNA monitoring for prognosticating whether patients with lung cancer will have favorable outcomes to standard-of-care therapies.^36,45 Our findings reveal that a large decrease in ctDNA corresponds with real-world outcomes for the NSCLC validity cohort, consistent with prior studies of ctDNA monitoring which have shown that ctDNA decrease or clearance is associated with improved PFS and OS for patients treated with immunotherapy in a variety of tumor types.^17,18,46 These studies often highlight the utility of quantifying ctDNA change, but our finding that a single on-treatment time point can stratify patient risk provides value in scenarios in which baseline ctDNA testing fails or in which pretreatment tissue profiling is preferred for therapy selection. In cases in which pretreatment ctDNA assessments are unavailable, on-treatment testing can identify patients who have achieved ctDNA clearance or who are persistently negative, both of which were associated with improved rwPFS and rwOS in this study.

In addition, we present the largest SCLC ctDNA monitoring study to our knowledge. As in our exploratory and NSCLC cohorts, on-treatment ctDNA level and a 90% decrease correlated with improved survival. SCLC is a tumor type in which the availability of tissue is often limited. These findings can help inform the design of clinical trials on utilizing ctDNA monitoring for high-risk populations, response to therapeutic agents, without subjecting patients to multiple biopsies, or in cases in which radiography is challenging. Notably, the difference in real-world response rates for patients with detectable ctDNA or less than 90% decrease was not statistically significant compared with those with undetectable ctDNA or 90% or higher decrease, respectively (Supplementary Tables 2 and 7), though the sample size was small. This may be due to treatment response typically being robust, as most patients would achieve a radiographic response regardless of ctDNA reduction. Still, the numerical differences in response rate were much smaller than what was observed in the NSCLC cohort. When ctDNA decrease and best real-world response were considered in conjunction, patients with both a molecular and imaging response had the best outcomes, particularly in SCLC (Fig. 5C). This suggests that in SCLC, in which initial tumor response (i.e., complete or partial radiographic response) is common, but one-year OS rate is low,^47,48 ctDNA monitoring may be a valuable complement to traditional imaging rather than relying on only imaging for response assessment. Thus, there may be a role in assessing MR in patients with radiographic response to select patients at high risk of relapse who can benefit from further maintenance therapy.

We saw higher rates of MR among patients with a documented rwR across all therapy types, but trends differed by therapy class. In patients treated with chemotherapy with or without ICIs, most patients experienced a decrease of 50% in ctDNA regardless of tumor response, whereas only 4/17 nonresponders receiving ICI monotherapy or targeted therapy had a ctDNA decrease of less than 50%. This is consistent with prior work, which found that early patterns of ctDNA clearance in patients treated with carboplatin with nab-paclitaxel compared with those in patients treated with atezolizumab in combination with the chemotherapy doublet are similar and that a higher cutoff for MR was needed to identify long term responders to the patients treated with chemotherapy alone.²⁵ We hypothesize that cytotoxic agents often used in combination with immunotherapy may cause a short-term decrease in ctDNA levels that is not correlated with tumor response, potentially confounding assessments of ctDNA reduction due to immunotherapy.

Questions remain surrounding the optimal cutoff for defining MR and when to collect on-treatment plasma. On the basis of previous studies,^40,49 we also explored definitions of MR as either a 50% or higher decrease or a 90% or higher decrease, but found no significant differences in outcomes between the two cutoffs. A 50% decrease was a more sensitive cutoff with regard to selecting patients with rwR, but this came at the expense of specificity (Supplementary Table 7). We assessed a broad range for the timing of the on-treatment blood draw, mirroring real-world practice. Although additional studies are needed to further explore the impact on the timing of specimen collection, these results support the utility of ctDNA monitoring in a real-world setting in which specimen collection will not always coincide with plasma collection intervals utilized in clinical trials.

The dynamics of specific VAFs can be discordant with ctDNA TF changes if the variant is derived from CH, which is common in genes such as DNMT3A. This highlights the need to properly contextualize genomic findings to discern tumor signals from germline or CH signals when analyzing results from tissue-agnostic monitoring. The quantification of ctDNA in this study did not rely on tissue or peripheral blood mononuclear cell sequencing. Although several studies have reported improved specificity when implementing peripheral blood mononuclear cell–based filtering, we previously reported a specificity higher than 98% for identifying CH using an algorithmic approach.⁴⁵ Furthermore, the relationship between ctDNA TF and VAF change in several oncogenes supports the accuracy of this algorithmic approach. We note that the availability of tissue results from only one-quarter of patients is indicative of the challenges inherent with a tissue-informed monitoring approach, particularly in lung cancer, in which tissue is limited.

We note several limitations of the study. First, rwR assessments used here are not directly analogous to the Response Evaluation Criteria in Solid Tumors, and, although previous studies have shown the validity of real-world end points, caution should be used when comparing end points such as ORR and PFS from the Response Evaluation Criteria in Solid Tumors–based studies to results presented here. Second, the timing of on-treatment ctDNA assessment was heterogeneous; the study was limited to patients with specimen collection six to 15 weeks, and risk set adjustment was performed to account for delayed entry. Further studies are needed to understand the optimal timing for ctDNA assessment by disease and therapy class. Third, the study design included patients with ctDNA assessment either on treatment or at progression. Clinical progression is often associated with increased ctDNA levels. When assessing MR, using ctDNA results from blood collected at clinical progression, although the minority of the patients (n = 53; 11% overall) in this study, may therefore lead to higher rates of ctDNA positivity and lower rates of MR than in studies in which plasma is collected during early treatment cycles. Moreover, although therapy classes were analyzed together, a wide variety of distinct regimens were used across variable lines of treatment, varying by dosing, frequency, and specific therapies among other variables. Additional follow up is needed to understand implications for ctDNA monitoring by line of therapy.

In conclusion, tissue-agnostic monitoring is feasible in patients with lung cancer across a wide range of therapy classes, including but not limited to standard of care immunotherapy regimens. Both on-treatment ctDNA detection and MR are strongly associated with patient outcomes in a real-world setting. This study indicates that on-treatment ctDNA dynamics differ between SCLC and NSCLC and by therapy class. Further studies are needed to identify the optimal definition of MR by disease and therapy class.

CRediT Authorship Contribution Statement

Anne C. Chiang: Conceptualization, Resources, Writing - original draft; and Writing - review & editing.

Russell W. Madison: Conceptualization, Supervision, Methodology, Formal analysis, Writing - original draft, Writing - review & editing, Software, Visualization.

Zoe June Assaf: Conceptualization, Methodology, Writing - original draft, Writing - review & editing.

Alexander Fine: Conceptualization, Data curation, Methodology, Writing - original draft, Writing - review & editing.

Yi Cao: Methodology, Writing - review & editing.

Ole Gjoerup: Writing - review & editing.

Yanmei Huang: Software, Writing - review & editing.

Dexter X. Jin: Software, Data curation, Writing - review & editing.

Jason Hughes: Software, Writing - review & editing.

Vladan Antic: Writing - review & editing.

Amanda Young: Writing - review & editing.

David Fabrizio: Writing - review & editing.

David Shames: Writing - review & editing.

Sophia Maund: Conceptualization, Methodology, Project administration, Writing - review & editing.

Alexia Exarchos: Project administration, Methodology, Writing - review & editing.

Shailendra Lakhanpal: Resources, Writing - review & editing.

Richard Zuniga: Resources, Writing - review & editing.

Lincoln W. Pasquina: Conceptualization, Supervision, Methodology, Writing - original draft, Writing - review & editing.

Katja Schulze: Conceptualization, Supervision, Methodology, Writing - original draft, Writing - review & editing.

Disclosure

Dr. Chiang is on the advisory boards or holds a consulting role at Amgen, AZ, Daichi, Fosun, GNE, Janssen, Jazz, Zai Labs; and is a research PI for AbbVie, Amgen, AstraZeneca, Bristol-Myers Squibb, GNE, Zai Labs. Mr. Madison, Dr. Fine, Dr. Gjoerup, Dr. Huang, Dr. Jin, Dr. Hughes, Dr. Young, Mr. Fabrizio, and Dr. Pasquina are employees of Foundation Medicine, Inc., a wholly owned subsidiary of Roche, with Roche stock ownership. Dr. Assaf, Dr. Cao, Dr. Shames, Dr. Maund, Ms. Exarchos, and Dr. Schulze are employees of Genentech, Inc. and own stock in Roche Holding AG. Dr. Antic is an employee of Hoffman-La Roche and owns stock in Roche Holding AG. Dr. Shames was an employee of Genentech, Inc. at the time of authorship. Dr. Zuniga holds consulting or advisory roles at Bristol Myers Squibb Foundation and Mirati Therapeutics. Dr. Lakhanpal declares no conflict of interest.