Circulating tumor DNA clearance predicts prognosis across treatment regimen in a large real-world longitudinally monitored advanced non-small cell lung cancer cohort.

BACKGROUND: Although growth advantage of certain clones would ultimately translate into a clinically visible disease progression, radiological imaging does not reflect clonal evolution at molecular level. Circulating tumor DNA (ctDNA), validated as a tool for mutation detection in lung cancer, could reflect dynamic molecular changes. We evaluated the utility of ctDNA as a predictive and a prognostic marker in disease monitoring of advanced non-small cell lung cancer (NSCLC) patients.
METHODS: This is a multicenter prospective cohort study. We performed capture-based ultra-deep sequencing on longitudinal plasma samples utilizing a panel consisting of 168 NSCLC-related genes on 949 advanced NSCLC patients with driver mutations to monitor treatment responses and disease progression. The correlations between ctDNA and progression-free survival (PFS)/overall survival (OS) were performed on 248 patients undergoing various treatments with the minimum of 2 ctDNA tests.
RESULTS: The results of this study revealed that higher ctDNA abundance (P=0.012) and mutation count (P=8.5×10-4) at baseline are associated with shorter OS. We also found that patients with ctDNA clearance, not just driver mutation clearance, at any point during the course of treatment were associated with longer PFS (P=2.2×10-16, HR 0.28) and OS (P=4.5×10-6, HR 0.19) regardless of type of treatment and evaluation schedule.
CONCLUSIONS: This prospective real-world study shows that ctDNA clearance during treatment may serve as predictive and prognostic marker across a wide spectrum of treatment regimens. 2020 Translational Lung Cancer Research. All rights reserved.

Introduction

The treatment of patients with advanced non-small cell lung cancer (NSCLC) has been revolutionized by the development of therapies targeting specific genetic alterations. The characterization of NSCLC into subtypes according to their genetic alterations has significantly improved the efficacy of targeted therapies and disease outcomes in subgroup of patients (1-4). However, their efficacies are compromised by development of resistance mechanisms (namely clonal evolution), which inevitably emerge in all patients with a median progression-free survival (PFS) ranging from a few months to a year (5-7). Response assessment primarily relies on imaging modalities, which may not reflect clonal evolution at molecular level (8). Therefore, there is a compelling need to develop improved modalities for monitoring clonal evolution.

The genomic profile of circulating tumor DNA (ctDNA), predominantly released by apoptosis and necrosis of cancer cells, has been shown to closely match that of tumor samples (9,10), and has now been validated as surrogate means for detecting mutations in NSCLC (11-13). For instance, plasma and tissue-based genotyping for EGFR T790M yielded equivalent clinical outcomes of osimertinib, thus supporting the use of plasma genotyping as an alternative diagnostic option (14). Much effort has been invested in exploring the potential of ctDNA in monitoring responses and assessing the emergence of drug resistance (15-17). Among patients undergoing epidermal growth factor receptor (EGFR) tyrosine kinase inhibitor (TKI) treatment, a reduction in the allelic fraction (AF) of EGFR mutation reflects sensitivity to these inhibitors (18). In addition, ctDNA has been instrumental in revealing novel resistance mechanisms, such as acquired EGFR C797S to osimertinib (5), MET Y1248H and D1246N to c-Met inhibitors, etc. (19).

Patients harboring the same mutation may exhibit marked differences in response to treatment (2). Circulating tumor DNA has been proposed as a noninvasive real-time biomarker to provide prognostic and predictive information for monitoring treatment (20-22). The prognostic value of ctDNA has been well-established in detecting minimal residual disease following surgery or treatment with curative intent, and is currently being explored in treatment responses of patients with advanced cancer (23-26). A recent study has shown that the detectable ctDNA at time of the diagnosis and identification of residual ctDNA at first evaluation were both associated with a poor prognosis (21). However, more work is needed to comprehensively examine its prognostic and predictive values in cohorts consisting of different treatment history.

In this prospective, real-world study, we performed capture-based ultra-deep targeted sequencing on longitudinal plasma samples to investigate the potential of ctDNA analysis in predicting clinical outcomes. We explored the genomic landscape of 1,336 Chinese patients with advanced NSCLC and subsequently focused on 248 of them with a minimum of 2 monitoring points for analyzing the predictive and prognostic value of ctDNA, as well as for investigating the dynamics of ctDNA upon pharmacological intervention by using a panel consisting of 168 NSCLC-related genes, covering 170KB of human genome.

Methods

Patient selection

From September 2015 to October 2016, advanced NSCLC (stage IIIB to IV) patients with specific mutations in at least one of the following genes EGFR, ALK, ROS 1, RET, KRAS, PIK3CA, ERBB2, MET and BRAF were enrolled. Their longitudinal plasma samples were collected at baseline and at various points throughout the ensuing treatment in multiple participating institutions. Detailed inclusion criteria were listed in supplemental methods. This study was approved by a central ethic committee at Nanjing General Hospital of Nanjing Command (2016NZKY-003-02). All other centers were covered by this protocol except for First Affiliated Hospital of Guangzhou Medical University (IRB2016-26) and Tianjin Medical School Affiliated General Hospital (IRB2016-050-01). All patients gave informed consent to participate in the study and gave permission for use of their peripheral blood.

Next generation sequencing (NGS) library preparation and capture-based targeted DNA sequencing

Fragments of size 200–400 bp were selected by AMPure beads (Agencourt AMPure XP Kit), followed by hybridization with capture probe baits, hybrid selection with magnetic beads and PCR amplification. Indexed samples were sequenced on Nextseq500 sequencer (Illumina, Inc., USA) with pair-end reads. An average depth of 11,816x was reached.

Statistical analysis

All statistical tests were conducted in R (version 3.3.1), using two-sided tests, unless otherwise specified. For patient characteristics, the differences in distribution of continuous and categorical variables across groups were assessed using Wilcoxon and Fisher exact tests, respectively. Survival tests were conducted using log-rank tests or Cox regression models when a co-variant was included.

Results

Patient demographics and study design

Within the screened population, 949 (71.03%) harbored driver mutations, 245 (18.34%) had no mutation detected, whilst the remaining 142 (10.63%) patients had non-driver mutations. Approximately 16% (n=207) patients were treatment-naïve; 71% (n=949) were previously treated and the remaining 13.5% (n=181) had no treatment history information available. Thirty-one percent of patients (n=410) had one line of previous treatment; 18.2% (n=244) had two lines; 11.1% (n=149) had three lines and the remaining 10.9% (n=146) had more than three lines of treatment (). The median follow-up time for patients enrolled in this study was 322 days (25–75%: 258–426 days). The median interval for ctDNA analysis was 95 days (25–75%: 82–120 days). shows detailed treatment history (outer ring) and treatment information during this study (inner ring). Among the 949 patients harboring driver mutations at baseline assessment, 376 patients received matched targeted therapy (MTT) according to sequencing results. Limited drug accessibility was the primary reason responsible for patients with driver mutation but unable to receive MTT. In addition, a significant number of patients had only EGFR sensitizing mutation upon progression on 1st generation EGFR-TKI, thus undergoing chemotherapy subsequently. A detailed view of their treatment prior to and during our study is shown in . Detailed survival analysis was performed on 248 patients (longitudinal cohort), with 2 or more evaluation time points beyond the baseline. A total of 280 patients had 2 or more follow-up tests and 32 of them were excluded due to various reasons. The selection of patients enrolled in the follow-up cohort is shown in . The remaining patients, with either baseline assessment or only one time follow-up subsequent to baseline, had limited information for detailed survival analysis. Therefore, they were excluded in such analyses.

Figure 1

Overview of our cohorts. (A) Schematic diagram delineates the presence or absence of driver mutations, treatment lines, follow-up time and number of ctDNA performed during the study. We screened 1,336 patients (screened cohort) to arrive at 949 patients with driver mutations to enroll in our study (enrolled cohort). Survival analyses were performed on 248 patients with 2 or more follow-up tests (longitudinal cohort). A total of 280 patients had 2 or more follow-up tests and 32 of them were excluded due to listed reasons. (B) This diagram illustrates the treatment history and treatment used in our study of the screened cohort. The outer ring represents treatment history and inner ring represents treatment used in our study. Different colors refer to different treatments.

Figure S1

Treatment history of patients treated with matched targeted therapy. This diagram illustrates the treatment history of patients who were treated with matched targeted therapy in our study. The outer ring represents treatment history and inner ring represents treatment used in our study. Different colors refer to different treatments.

We first compared and contrasted baseline clinical parameters, including gender, age, smoking history, histology, stage, treatment history and metastatic sites between longitudinal and screened cohorts. Our data demonstrated that the two cohorts were similar in most of the parameters, except for gender, presence of bone metastasis and EGFR mutation status (). The longitudinal cohort had female predominance, a larger number of bone metastasis and EGFR mutations. These associations are due in part to EGFR being the most frequent driver mutation in Asian NSCLC patients, in particular female patients. However, such differences do not skew analyses performed in this study.

Table S1

Comparison of screened cohort and selected cohort

Characteristics	Screened cohort			Selected cohort		P value
	n=1,336	%		n=248	%
Gender						0.032
Male	648	51.2		108	44
Female	652	48.8		140	56
Stage						0.243
IIIB	70	6.8		11	17
IV	965	93.2		235	83
Smoking history						0.375
Y	135	20.4		11	5
N	526	79.6		235	95
Histological type						0.882
Adenocarcinoma	928	91.2		218	92.4
Squ carcinoma	43	4.2		8	3.4
Adenosqu carcinoma	36	3.5		9	3.8
Combined SCLC	10	1		1	0.4
Treatment history						0.099
Treatment-naive	196	17.6		54	22.2
Previously treated	919	82.4		189	78.8
Metastatic sites
Bone	465	34.8		140	56.5	0.003
Liver	175	13.1		54	21.8	0.117
Brain	287	21.5		64	25.8	0.428
Driver mutations
ALK						0.426
Mutation	64	4.8		15	6
WT	1,272	95.2		233	94
BRAF						0.095
Mutation	18	1.3		0	0
WT	1,318	98.7		248	100
EGFR						1.22E-15
Mutation	700	52.4		196	79
WT	636	47.6		52	21
ERBB2						0.119
Mutation	30	2.2		10	4
WT	1,306	97.8		238	96
KRAS						0.22
Mutation	76	5.7		9	3.6
WT	1,260	94.3		239	96.4
MET						0.274
Mutation	11	0.8		4	1.6
WT	1,325	99.2		244	98.4
PIK3CA						0.853
Mutation	49	3.7		8	3.2
WT	1,287	96.3		240	96.8
RET						0.754
Mutation	17	1.3		2	0.8
WT	1,319	98.7		246	99.2
ROS1						0.704
Mutation	11	0.8		1	0.4
WT	1,325	99.2		247	99.6

Squ, squamous; SCLC, small cell lung cancer; WT, wild type.

Landscape of baseline mutation

We performed capture-based ultra-deep targeted sequencing on all baseline plasma samples using a panel consisting of 168 genes, spanning 170KB of human genome. The design and validation of this panel, earlier described by Mao et al. (9), achieved 95% and 87% by-variant sensitivity for identifying mutations from matched tissue and plasma samples, respectively, excluding copy number variations (CNVs) (9). DNA obtained from white blood cells (WBCs) was used as a reference to sort out germline mutations. Overall, an average of 11,816× sequencing depth was achieved.

At baseline, we identified 3,503 aberrations spanning 132 genes, including 2,204 single-nucleotide variants (SNVs), 693 insertions or deletions (Indels), 412 copy-number amplifications (CNAs), 80 copy number deletions, and 114 translocations. Approximately 18% patients (245/1,336) had no mutations detected from this panel. EGFR was the most frequently mutated gene, followed by TP53, occurring in 55% and 41% of patients, respectively. Among all genetic aberrations identified, well-established NSCLC driver mutations, including EGFR, KRAS, BRAF, ERBB2, ALK, RET and ROS1, comprised 46.9% of all variants. The overview of mutation spectrum is shown in .

Figure 2

Mutations identified in baseline plasma samples. (A) OncoPrint of mutations identified at baseline of the screened cohort. Different colors denote different types of mutations. Top bar represents the number of mutations a patient carries; side bar represents the number of patients carry a certain mutation. Bottom bars provide information regarding histology, gender and treatment history. (B) Clinical characteristics (M stage, presence of bone metastasis, presence of liver metastasis and number of organs with lesions) associated with maxAF and cfDNA. Pearson or t-test correlation test was applied for continuous variables or binary variables, respectively. Boxplots of both variables over the dichotomized clinical features are shown.

We then investigated the clinical relevance of baseline maximum allelic fraction (maxAF) and total cell-free DNA (cfDNA). MaxAF was defined as the maximum allelic fraction among all somatic mutations identified in a plasma sample. Higher maxAF and cfDNA were associated with more advanced M stage, a higher likelihood of bone/liver metastasis and more organs with secondary lesions (). Interestingly, maxAF showed a more significant correlation with all clinical features than the amount of cfDNA.

Overall survival (OS) is correlated with baseline ctDNA abundance and mutation load

We performed detailed analysis on the longitudinal cohort to assess the predictive and prognostic value of ctDNA. We first investigated the correlation between OS and baseline parameters, including ctDNA abundance and mutation load. Previous studies exploring the correlation between mutation load at baseline and OS provided controversial or inconsistent results (27,28). Our data revealed an inverse correlation between baseline ctDNA amount (expressed as the product of maxAF and total amount of cfDNA and OS (P=0.012). The mutation count was also inversely correlated with OS, independent of baseline ctDNA amount (P=8.5×10−4) (). We then derived a molecular signature for OS prediction using multivariate stepwise regression, starting from 6 genes that individually associated with OS: CDKN2A, EGFR, KEAP1, KRAS, MET and POM121L12. The final molecular signature consisted of KEAP1, KRAS and MET. Patients with no mutations in these genes had longer OS (P<0.0001) ().

Figure 3

Correlation between baseline characteristics and overall survival. (A) ctDNA. (B) Mutation count. (C) A signature consisting of KEAP1, KRAS and MET can predict OS. Patients with no mutation in the above 3 genes have a longer OS than patients with mutation in any one of the above 3 genes. *, denotes P value derived from cox regression model.

ctDNA clearance predicts longer PFS and OS

In clinical settings, treatment response is typically monitored on a regular interval by radiological imaging, which will not mirror clonal evolution. We analyzed the potential use of ctDNA as a surrogate marker for monitoring treatment response in our longitudinal cohort, which had at least 2 ctDNA tests. After a median follow-up of 157 days, disease progression occurred in 166 (66.9%) patients. During the course of treatment, 123 patients treated with either MTT or chemotherapy had a minimum of one time ctDNA clearance occurring from 1 to 15 months after commencement of treatment, with a median PFS of 8.6 months. CtDNA clearance is defined as lack of detectable mutation from this panel covering 168 lung cancer-related genes, with an average sequencing depth of 11,816× and 0.2% detection limit. Fifty patients achieved partial response (PR), 67 had stable disease (SD) and 3 had progressive disease (PD) as the best response, thus yielding to an overall response rate (ORR) of 41.7% and a disease control rate (DCR) of 97.5%. Up to June 25, 2017, the median OS of this group has not been reached. Conversely, 125 patients with consistent detectable ctDNA throughout the course of treatment had a median PFS of 4.1 months and a median OS of 16.7 months. Among them, 14 achieved PR, 64 had SD and 38 had PD as the best response, thus yielding to an ORR of 12.1% and a DCR of 67.2%. Taken together, this data reveals that patients with a minimum of one time ctDNA clearance experienced longer PFS (P=2.2×10−16; HR 0.28) and longer OS (P=4.5×10−6, HR 0.19) independent of baseline ctDNA amount, regardless of type of treatment and time of evaluation (). The baseline clinical parameters including gender, smoking history, stage, treatment history of patients with the minimum of one time ctDNA clearance and those with consistently detectable ctDNA were comparable, except for gender. More male patients experienced ctDNA clearance (). Furthermore, patients with the minimum of one time ctDNA clearance had better ORR (P=3.9×10−7) and DCR (P=1.4×10−10) compared to those with detectable ctDNA throughout the course of treatment. A similar trend was observed in patients undergoing MTT (), but not in those treated with chemotherapy (). Overall, ctDNA clearance was a significant predictor of PFS (P=0.022) but not OS (P=0.22) in chemotherapy-treated patients after adjustment for baseline ctDNA amount. Collectively, this data shows ctDNA may be a valuable real-time biomarker for monitoring therapeutic response, whilst its clearance at any point of treatment may efficiently predict treatment response. This aspect actually mirrors clonal response, thus allowing uncovering the biological nature beyond the clinical response.

Figure 4

Predictive and prognostic value of ctDNA clearance during the course of treatment. (A) Kaplan-Meier curves for PFS and OS in patients with a minimum of one time ctDNA clearance vs patients with consistent detectable ctDNA throughout the course of treatment. (B) patients treated with MTT. (C) ROC curve for changes in maxAF during the course of treatment. A reduction of maxAF to zero is the optimal cutoff with an AUC of 75%. (D) Kaplan-Meier curves for PFS in patients with driver mutation clearance, all mutation clearance and patients with the presence of both driver and other mutations throughout the course of treatment.

Table S2

Comparison of patients without ctDNA clearance and patients with ctDNA clearance

Characteristics	No ctDNA clearance			ctDNA clearance (n≥1)		P value
	n=125	%		n=123	%
Gender						0.015
Male	64	51.2		44	64.2
Female	61	48.8		79	35.8
Stage						0.767
IIIB	5	4		6	4.9
IV	119	96		116	95.1
Smoking history						1
Y	14	16.5		13	17.6
N	71	83.5		61	82.4
Histological type
Adenocarcinoma	108	91.5		110	93.2	0.882
Squ carcinoma	5	4.2		3	2.5
Adenosqu carcinoma	4	3.4		5	4.2
Combined SCLC	1	0.8		0	0
Treatment history						0.122
Treatment-naive	33	26.6		21	17.1
Previously treated	91	73.4		102	82.9
Metastatic sites
Bone	74	59.7		66	55.9	0.603
Liver	32	25.8		22	18.6	0.217
Brain	33	26.6		31	26.3	1
Driver mutations
ALK						1
Mutation	9	7.2		6	4.9
WT	116	92.8		117	95.1
EGFR						0.596
Mutation	95	76		101	82.1
WT	30	24		22	17.9
ERBB2						0.749
Mutation	6	4.8		4	3.3
WT	119	95.2		119	96.7
KRAS						0.5
Mutation	6	4.8		3	2.4
WT	119	95.2		120	97.6
MET						0.37
Mutation	1	0.8		3	2.4
WT	124	99.2		120	97.6
PIK3CA						0.722
Mutation	5	4		3	2.4
WT	120	96		120	97.6
RET						0.245
Mutation	0	0		2	1.6
WT	125	100		121	98.4
ROS1						0.496
Mutation	0	0		1	0.8
WT	125	100		122	99.2

Squ, squamous; SCLC, small cell lung cancer; WT, wild type.

Figure S2

Predictive and prognostic value of ctDNA clearance during the course of treatment. Kaplan-Meier curves for PFS and OS in chemotherapy-treated patients with a minimum of one time ctDNA clearance vs patients with consistent detectable ctDNA throughout the treatment.

Previous studies showed that decreasing ctDNA levels during treatment may be associated with favorable therapeutic efficacy (21). We hence investigated whether or not ctDNA clearance or certain degree of ctDNA reduction reflected by maxAF would better correlate with treatment efficacy. To derive a binary classifier which can differentiate the population according to treatment efficacy, we performed receiver operating curve (ROC) analysis of changes in maxAF during the course of treatment. A decrease of maxAF to zero was identified as the optimal cutoff, characterized by area under curve (AUC) of 0.75 (95% CI: 0.691–0.802) (). Change in maxAF was defined as the ratio of smallest maxAF detectable in follow-up evaluations and baseline. Therefore, ctDNA clearance, but not ctDNA decrease in response to pharmacological interventions, could be identified as a predictive marker.

We then evaluated whether or not the clearance of driver mutation may provide similar predictive power as ctDNA clearance for predicting PFS. We compared PFS among 3 groups of patients, with all mutation clearance, only driver mutation clearance, and with presence of driver mutations. No difference in PFS was observed in patients with only driver mutation clearance and those with driver mutation, thus suggesting that monitoring only driver mutations will not efficiently predict PFS ().

Discussion

Therapeutic response in cancer is conventionally assessed with imaging techniques, which are however unable to identify clonal dynamics and evolutionary changes during therapeutic management. Recent studies showed that ctDNA may be a reliable tool for real-time tracking of molecular dynamics and hence for predicting treatment response defined by residual disease (21,22). In this study, we investigated the diagnostic value of ctDNA by performing capture-based ultra-deep sequencing on longitudinal plasma samples obtained at baseline and multiple efficacy evaluation time points from 248 patients with advanced NSCLC. Our real-world study, comprising both treatment-naïve and previously treated patients, shows that ctDNA may be a valuable real-time biomarker for monitoring therapeutic response, whilst ctDNA clearance at any point of treatment may also be used for efficiently predicting treatment benefits. CtDNA clearance was defined as no detectable mutations using this panel that covers 168 lung cancer-related genes, characterized by average sequencing depth of 11,816× and 0.2% detection limit. Albeit we cannot rule out the possibility that patients with “ctDNA clearance” had mutations with AFs below the detection limit (i.e., 0.2%), patients with the minimum of one time ctDNA clearance during the course of treatment had statistically significant longer PFS and OS compared to those with detectable ctDNA during therapy.

The prognostic value of ctDNA at first assessment (baseline) has been previously described in a prospective study comprising only newly diagnosed patients undergoing first-line treatment (21). Our study, consisting of a heterogeneous population and diverse evaluation schedules, not only confirmed the finding of this earlier investigation, but also extended the significance of ctDNA analysis for predicting treatment benefits in all patients, regardless of treatment history and time of evaluation. We also demonstrated that the assessment of ctDNA clearance may generate valuable clinical benefits, namely longer PFS and OS. However, our findings warrant additional studies aimed to define the value of ctDNA clearance as surrogate endpoint of therapeutic efficacy and as a risk stratification factor, especially for differentiating poor from favorable patient outcomes.

The prognostic and predictive value of ctDNA level before treatment remains a controversial issue. Many studies showed that increased ctDNA level at baseline is associated with unfavorable PFS and OS, while others failed to find a significant correlation with clinical outcomes (21,29,30). We observed a significant inverse correlation between baseline ctDNA amount and OS, while we also identified a molecular signature predictive of OS. A recent study has identified new determinants of ctDNA in NSCLC, including necrosis degree, lymph node involvement, lymphovascular invasion, pathological tumor size, Ki-67 labelling indices and tumor histology (31). Therefore, large cohort studies controlling for these factors would be necessary to accurately define the prognostic and predictive value of the baseline ctDNA assessment. There are some potential limitations associated with this study, including heterogeneity in treatments and time points in ctDNA evaluation.

Conclusions

To the best of our knowledge, this is the largest real-world study consisting of Chinese patients with advanced NSCLC to evaluate the value of ctDNA in monitoring treatment response. Taken together, our study showed the predictive and prognostic value of ctDNA clearance during treatment in a heterogeneous population with diverse treatment regiments and evaluation schedules.

The article’s supplementary files as