Prognostic and Predictive Impact of Circulating Tumor DNA in Patients with Advanced Cancers Treated with Immune Checkpoint Blockade

Abstract

The utility of circulating tumor DNA (ctDNA) as a biomarker in patients with advanced cancers receiving immunotherapy is uncertain. We therefore analyzed pretreatment (n = 978) and on-treatment (n = 171) ctDNA samples across 16 advanced-stage tumor types from three phase I/II trials of durvalumab (± the anti-CTLA4 therapy tremelimumab). Higher pretreatment variant allele frequencies (VAF) were associated with poorer overall survival (OS) and other known prognostic factors, but not objective response, suggesting a prognostic role for patient outcomes. On-treatment reductions in VAF and lower on-treatment VAF were independently associated with longer progression-free survival and OS and increased objective response rate, but not prognostic variables, suggesting that on-treatment ctDNA dynamics are predictive of benefit from immune checkpoint blockade. Accordingly, we propose a concept of "molecular response" using ctDNA, incorporating both pretreatment and on-treatment VAF, that predicted long-term survival similarly to initial radiologic response while also permitting early differentiation of responders among patients with initially radiologically stable disease. SIGNIFICANCE: In a pan-cancer analysis of immune checkpoint blockade, pretreatment ctDNA levels appeared prognostic and on-treatment dynamics predictive. A "molecular response" metric identified long-term responders and adjudicated benefit among patients with initially radiologically stable disease. Changes in ctDNA may be more dynamic than radiographic changes and could complement existing trial endpoints.This article is highlighted in the In This Issue feature, p. 1775.

Trial registration: ClinicalTrials.gov NCT01693562 NCT02087423 NCT02261220.

©2020 American Association for Cancer Research.

Figures

Figure 1.

A, Somatic alteration detection rate by tumor type. Detection rate varied from 21.4% for patients with GBM to >95% for patients with SCLC and nasopharyngeal carcinoma. The error bars denote the 95% CI of detection rate for each tumor type given the sample sizes. B, Kaplan–Meier analysis of OS in the discovery cohort stratified by ≥median versus <median pretreatment VAF. For OS, unadjusted HR, 0.58; 95% CI, 0.49–0.69; P < 0.0001 and adjusted HR, 0.58; 95% CI, 0.48–0.7; P < 0.0001. C, Pretreatment VAF stratified by objective response (complete and partial responses vs. SD and progressive disease) in the discovery and validation cohorts. Pretreatment VAF was not significantly associated with response in any cohort. The horizontal bar represents the mean, the box represents the 25th and 75th percentiles, and whiskers ± 1.5 × interquartile range. Kaplan–Meier analysis of OS in the validation cohorts. D, Pretreatment VAF was not significantly associated with OS in ATLANTIC (HR, 0.58; 95% CI, 0.3–1.16; P = 0.12). E, Pretreatment VAF was significantly associated with OS in Study 10 (HR, 0.42; 95% CI, 0.24–0.74; P = 0.0019). F, Forest plot for OS by tumor type stratified by median pretreatment VAF using pooled data from the discovery cohort and both validation cohorts. CR, complete response; HNSCC, head and neck squamous cell carcinoma; HPV, human papillomavirus; PD, progressive disease; PR, partial response; TNBC, triple-negative breast cancer; UC, urothelial cancer.

Figure 2.

A, Forest plot of PFS and OS by on-treatment VAF. The best OS was observed in patients with complete ctDNA clearance on treatment. B, RECIST responders and nonresponders had significantly different on-treatment VAF (P < 0.0001). C and D, Kaplan–Meier analysis of PFS (C) and OS (D) in the discovery cohort stratified by delta-VAF and on-treatment VAF. Significant differences were observed among groups (P < 0.0001). §§, P < 0.00001.

Figure 3.

Development of a definition of molecular response with contributions from pretreatment VAF and on-treatment VAF. A, ROC curves of molecular response and on-treatment VAF to predict the best response. B, ORR among patients with and without response as defined by either the ratio (molecular response) or on-treatment VAF. C, Forest plot for PFS and OS based on molecular response or the first RECIST response in each study.

Figure 4.

A, Number of molecular responders and molecular nonresponders (per ratio approach) with eventual RECIST response. A higher percentage of eventual radiologic responders was initial molecular responders (19/25, 76%) than the percentage of eventual radiologic nonresponders who were initial molecular responders (12/49, 24%; P < 0.0001, Fisher test). B, Swimmer plot of 14 patients with initial SD who were molecular responders and eventually had a radiologic response. The median time to radiologic response was 114 days from starting of the therapy, a median of 59 days later than the ctDNA assessment. C, Tumor shrinkage from baseline was significantly greater in molecular responders than molecular nonresponders (per ratio approach; P < 0.0001). The horizontal bar represents the mean, box represents the 25th and 75th percentiles, and whiskers ± 1.5 × interquartile range. D and E, Kaplan–Meier analysis of PFS (D) and OS (E) of patients who had RECIST and molecular response assessments stratified by molecular response (per ratio approach). PFS and OS were significantly longer in molecular responders than in molecular nonresponders (HR, 0.31; 95% CI, 0.17–0.56; P = 0.0001 and HR, 0.36; 95% CI, 0.17–0.79; P = 0.008, respectively). ††, P < 0.0001.