Our research center is one of the sub-centers of the RATIONALE-302(17) study (ClinicalTrials.gov identifier no. NCT03430843), of which 20 subjects were screened and 15 were enrolled, of which 12 patients were eligible for this exploratory study at Zhejiang Cancer Hospital (Hangzhou, China) between March 2018 and September 2020, whose tumor tissue or plasma was available for NGS. Eligible patients were aged 18-73 years and previously treated with systematic chemotherapy for stage IV ESCC, who were being treated with tislelizumab immuno-monotherapy 200 mg intravenous drip every 3 weeks until disease progression. Specific inclusion and exclusion criteria were those of the RATIONALE-302 study (17). This study was approved by the Institutional Review Board of Zhejiang Cancer Hospital (Hangzhou, China) and all patients provided written informed consent. The study was conducted in accordance with the established ethical standards of the institutional and/or national research committee, as outlined in the Declaration of Helsinki.

Patients underwent CT imaging at a follow-up visit at every 6-8 weeks during immunotherapy and chest and abdominal CT scans were evaluated according to Response Evaluation Criteria in Solid Tumours (RECIST 1.1) (18), by a radiologist blinded to circulating cell-free DNA (cfDNA) data. Progression-free survival (PFS) was defined as the interval between treatment initiation and the date of disease progression or death from any cause, whichever occurred earlier. Overall survival (OS) was defined as the time interval from treatment initiation to death from any cause. A censor date of June 30, 2021 was applied if no endpoint was met.

Fresh ESCC tumor samples from patients undergoing esophagogastroscopy were obtained from 12 patients before immunotherapy. A total of 35 blood samples for ctDNA were obtained at baseline, when patients achieved a status of partial response (PR) or stable disease (SD), or after progressive disease (PD) from March 2018 through to September 2020. Tumor samples were subjected to pathological diagnosis, graded according to the World Health Organization system (19) and then prepared into formalin-fixed and paraffin-embedded (FFPE) blocks. Peripheral blood (8-10 ml) was collected from each patient at each time-point in EDTA-coated tubes (BD Biosciences), and plasma was separated within 2 h after peripheral blood collection. All samples were sent to the centralized testing center of Nanjing Geneseeq Technology, Inc. (Nanjing, China) for targeted NGS.

Genomic DNA from FFPE tissues was extracted using a QIAamp DNA FFPE Tissue Kit (Qiagen GmbH). DNA from white blood cells (WBC) was extracted with the DNeasy Blood & Tissue Kit (Qiagen GmbH) and sequenced as normal controls to distinguish germline variations. Plasma-derived cfDNA was extracted using the QIAamp Circulating Nucleic Acid Kit (Qiagen GmbH). DNA was quantified with a Qubit 3.0 (Thermo Fisher Scientific, Inc.) and the quality was assessed using a Nanodrop 2000 (Thermo Fisher Scientific, Inc.). FFPE and WBC-derived DNA was sheared into 300-350 bp fragments using a Covaris M220 instrument (Covaris, Inc.). cfDNA or fragmented genomic DNA was prepared using a KAPA Hyper Prep Kit (KAPA Biosystems) as previously described (20). In brief, the DNA was subjected to end repair, A-tailing and adaptor ligation. DNA was then amplified and purified. A customized NGS panel containing 425 cancer-related genes was used for library enrichment. The captured libraries were amplified, purified and quantified. All the experimental procedures were performed at Nanjing Geneseeq Technology, Inc.

The prepared library was sequenced using a HiSeq4000 platform (Illumina, Inc.). The mean coverage depth was 150x µm for WBC controls and 800x µm for tissue samples. For cfDNA samples, the mean coverage sequencing depth was 5,000x. Base calling was performed on a bcl2fastq v2.16.0.10 (Illumina, Inc.) to generate sequence reads in the FASTQ format (Illumina 1.8+ encoding). Quality control was performed using the Trimmomatic software version 0.40 (https://github.com/usadellab/Trimmomatic). High-quality reads were aligned to the human genome (hg19; Genome Reference Consortium Human ref. 37) using Burrows-Wheeler Aligner 0.7.12 (https://github.com/lh3/bwa/tree/master/bwakit). The data were further processed using Picard 1.119 (https://github.com/broadinstitute/picard/releases/latest) and the Genome Analysis Toolkit (GATK; version 3.4.0; https://software.broadinstitute.org/gatk).

Single nucleotide variants (SNVs) and short insertions/deletions (indels) were identified using VarScan2 and HaplotypeCaller/UnifiedGenotyper in GATK, with a minimum mutant allele frequency cut-off of 0.5% for tissue, 0.3% for cfDNA and a minimum of three unique mutant reads with good quality scores. All SNVs/indels were annotated using ANNOVAR (default version; https://bioweb.pasteur.fr/packages/pack@annovar@24.10.2019) and manually checked using the Integrative Genomics Viewer version 2.10.3 (https://igv.org). Gene fusions/rearrangements were identified using FACTERA version 1.4.4 (https://factera.stanford.edu/download.php) and copy number gain or loss was analyzed using ADTEx version 2.0 (https://sourceforge.net/projects/adtex). The log2 ratio cutoff defined for copy number gain was 2.0 for tissue and 1.6 for cfDNA. A log2 ratio cutoff defined for copy number loss was 0.6 for all sample types. Tissue tumor mutation burden (tTMB) was calculated according to Fang's study (21). High-level tTMB (tTMB-H) was defined as ≥10 mutations/Mb (mut/Mb) and <10 mut/Mb was defined as tTMB-L.

A longitudinal ctDNA analysis was performed and the mutant molecules per milliliter of plasma (MMPM) were measured, which quantifies ctDNA for all variants. The overall MMPM for each sample was defined as the mean MMPM (mMMPM) across SNVs and indels, which was calculated as previously described (22,23). In brief, the mMMPM was calculated from the mean allele fraction (mAF) of variants by multiplying by the extracted mass (ng), dividing by the plasma volume (ml) and adjusting by a factor of 330 haploid human genome equivalents per ng:

∆mMMPM was defined as the percentage change of mMMPM in plasma at the first monitoring time-point after the first infusion of tislelizumab (ranging from cycles 2 to 6) since baseline for each patient:

Quantitative data are presented as the median (range) or n (%). Proportion comparisons between the groups were performed using Fisher's exact test. Kaplan-Meier curves were used to perform the survival analysis. Genes with mutations that were significantly associated with PFS and OS were screened using univariate Cox models. A two-sided P<0.05 was considered statistically significant for all tests unless indicated otherwise. All analyses were performed using R software version 3.6.3 (https://cran.r-project.org/bin/windows/base/old/3.6.3/).

Between March 2018 and September 2020, 15 patients who received tislelizumab as a second-line treatment for advanced ESCC in the RATIONALE-302 study (17) were recruited for the present study. Among them, 12 patients had endoscopic biopsy samples taken before tislelizumab immunotherapy, as well as plasma samples before and after every 2-3 (6-9 weeks) treatment cycle and disease progression. These 12 patients were included in the subsequent analysis. All 12 patients had interpretable data at the pretreatment time-point and 10 patients (83.3%) had interpretable data of NGS at the dynamic time-points and were used for analyses. The median age of the patients was 62.5 years (range, 46-73 years). There were 11 men (91.7%), nine current or former heavy drinkers (75.0%) with >1 l of liquor per day. The Eastern Cooperative Oncology Group performance status (24) of all the 12 patients was 0-1. Among these 12 patients, a total of four patients (33.3%) had undergone surgery and nine patients (75.0%) had received local radiotherapy, one of them had undergone both surgery and radiotherapy. Of the 12 patients, nine (75%) had evaluable lymph node metastases, six (50%) had lung metastases and two (16.7%) had liver metastases. All patients received tislelizumab as the second-line treatment. A total of six patients (50.0%) received ≥12 cycles of tislelizumab. The overall response rate was 33.3%. After a median follow-up of 19.3 months, eight patients (66.7%) exhibited confirmed tumor progression, of whom six (50.0%) died (Table I). The median PFS of the 12 patients treated with tislelizumab was 7.9 months (Fig. 1A) and the median OS was 17.6 months (Fig. 1B). Tumor evaluation was performed after every 2 injections of tislelizumab at 6 weeks of treatment for all patients, except for one patient with severe immune-related thrombocytopenia, who died quickly after only one injection of tislelizumab. This patient only underwent endoscopic biopsy and plasma samples were taken before treatment and were included in the following ctDNA analyses. Of the 12 patients with ESCC in the present study, almost all clinical characteristics (such as age, sex, treatments, cycles of treatment and number of metastatic organs) were not significantly associated with clinical response and outcome (data not shown), except for liver metastasis. The present results showed that patients with liver metastasis had a significantly shorter mPFS (1.4 vs. 11.7 months; P=0.037) (Fig. 1C) and a tendency of shorter OS (13.2 vs. 17.6 months; P=0.560) (Fig. 1D). Table I describes the population characteristics and clinical response to tislelizumab immunotherapy of the patients of this study.

Tumor tissues were available from all 12 patients and were subjected to targeted sequencing of 425 cancer-associated genes. Tumor mutation profiles included high-frequency alterations in tumor protein (TP)53 and cell cycle genes, such as fibroblast growth factor (FGF)19 and cyclin (CCN)D1 (Fig. 2A). According to the Kaplan-Meier survival analysis, the median OS for TSC complex subunit 2 (TSC2) and zinc finger protein 217 (ZNF217) gene mutation was significantly shorter than that for TSC2 [11.7 months vs. not reached (NR), P=0.004, Fig. 2B] or ZNF217 wide type (11.7 months vs. NR, P=0.022, Fig. 2C), while this trend was not significant for PFS. Furthermore, patients with FGF19 gene amplification had a slightly longer PFS (Fig. S1A) and OS (Figs. 2D and S1B) than patients with FGF19 gene wild-type, but the differences were not significant. However, when patients were stratified by the tTMB, patients with tTMB-H had longer PFS and OS than patients with tTMB-L, although no significant differences were obtained (PFS: 10.8 vs. 7.2 months, P=0.750; OS: 17.0 vs. 13.2 months, P=0.680) (Fig. S1C and D).

Subsequently, pretreatment ctDNA levels were analyzed in 12 patients for whom baseline plasma samples were available. Table SI shows the ctDNA results for all 35 plasma samples collected by NGS for each patient. ctDNA was detected in 7/12 (58.3%) patients at baseline. Patients with undetectable ctDNA at baseline showed prolonged OS compared to those with detected ctDNA (NR vs.13.2 months; P=0.017) and had a better tendency in PFS (18.2 vs. 5.5 months; P=0.078) (Fig. 3A and B). The mMMPM in these 10 patients at baseline was 91.68 (range: 0-462.72). ctDNA remained undetectable in two patients during the entire serial monitoring.

Three patterns of ctDNA molecular status were observed in the 10 patients who had plasma samples taken after tislelizumab treatment.

The first pattern, seen in two of the ctDNAs undetectable in the whole serial monitoring (Patients 11 and 12), was consistent with radiologic PR or SD for >1 year up to the end of follow-up time (Fig. 3C and D). In these two cases, tumor-specific variants were undetectable during serial monitoring of tislelizumab immunotherapy. The follow-up duration for these two patients were 12 and 17 months, respectively, and no disease progression had occurred at the last follow-up.

In the second pattern, ctDNA levels displayed a relatively obvious rise (defined as ∆mMMPN ≥20%) 2-6 cycles after therapeutic initiation. However, two patients with ∆mMMPN ≥20% showed no clinical benefit. One patient exhibited PD and another patient showed SD with a PFS of <4 months. As a representative case, ctDNA levels in patient 04 continued to rise from the time of initiation of tislelizumab immunotherapy, which is consistent with radiographically confirmed PD (Fig. 4A).

The last pattern, among the patients with ∆mMMPM <20% (n=8), showed a marked reduction or no relatively obvious rise in ctDNA on average at 3 cycles from treatment initiation. All patients showed clinical benefit with at least PR or SD, with a PFS of >6 months, except for one patient who had assessed PD. For instance, for patient 05, ctDNA-based molecular analyses showed a marked reduction in molecular response at cycle 2, which coincided with a radiologic partial response, and at the time of acquired resistance at 60 cycles, while radiographic analysis showed disease progression according to RECIST 1.1(18), consistent with rising ctDNA levels during tislelizumab immunotherapy (Fig. 4B).

Accordingly, ctDNA dynamic changes expressed as the ∆mMMPM more accurately predicted PFS and OS than tTMB (Fig. S1C and D). Patients with ∆mMMPN<20% had significantly longer PFS than patients with ∆mMMPN ≥20% (14.6 vs. 2.8 months; P=0.029) (Fig. 4C), although no significant differences were observed in OS (17.6 vs. 16.7 months; P=0.830) (Fig. 4D). To a certain extent, ∆mMMPN <20% was an independent predictor of PFS in patients receiving second-line tislelizumab monotherapy for advanced and metastatic ESCC.