Abstract

   Esophageal squamous cell carcinoma (ESCC) remains a major contributor
   to cancer-related mortality, with molecular residual disease (MRD)
   detection posing a significant challenge in post-surgical management.
   This study aimed to evaluate the effectiveness of circulating tumor DNA
   (ctDNA) in detecting MRD in patients with resectable ESCC undergoing
   radical surgery. A total of 62 primary tumor tissues, 108 preoperative,
   and 125 postoperative plasma samples were collected from 125 such ESCC
   patients who underwent radical surgery, and subjected to sequencing.
   Next-Generation Sequencing and ultra-high sensitivity Automated Triple
   Groom Sequencing panels were used to sequence genomic DNA from tumor
   tissues and ctDNA from plasma, respectively. ctDNA positive mutations
   included tumor-informed mutations and tumor-naïve mutations. Key
   findings revealed a high concordance rate of mutation detection between
   primary tumor tissue and preoperative plasma samples (91.11%,
   p = 0.62). Critically, the recurrence rate was higher in postoperative
   ctDNA-positive ESCCs than that in negative ones (66.67% (40/60) vs.
   21.54% (14/65), p < 0.001). And postoperative ctDNA-positivity was
   associated with poorer disease-free survival (DFS) (hazard ratio (HR):
   4.58, 95% CI: 2.65–7.92, p < 0.001) and overall survival (OS) (HR:
   5.39, 95% CI: 2.96–9.80, p < 0.001). Similar prognostic patterns were
   observed in patients with preoperative ctDNA-positivity (p = 0.014;
   p = 0.016; p = 0.071) and ctDNA-nonclearance (p < 0.001; p < 0.001;
   p < 0.001). Furthermore, in combination with postoperative ctDNA
   status, the Tumor-Node-Metastasis-Blood (TNMB) staging system was able
   to better distinguish patients with different prognoses compared with
   traditional TNM (p < 0.001). In conclusion, postoperative
   ctDNA-positivity emerges as a promising biomarker for detecting MRD in
   ESCC patients following surgical resection.

Supplementary Information

   The online version contains supplementary material available at
   10.1186/s43556-025-00310-6.

   Keywords: Esophageal squamous cell carcinoma, Circulating tumor DNA,
   Molecular residual disease, Biomarker, Circulating cell-free DNA

Introduction

   In 2022, esophageal cancer (EC) was the 11th most diagnosed cancer
   globally (511,000 cases) and the 7th deadliest (445,000 deaths)[[46]1].
   China is a country with a high incidence of EC, accounting for half of
   these cases and deaths[[47]2], predominantly esophageal squamous cell
   carcinoma (ESCC) (~ 90%). Surgery remains the cornerstone in treatment
   for resectable ESCCs, but over half of the patients face recurrence and
   metastasis post-surgery[[48]3], resulting in a low 5-year survival rate
   of 25%[[49]4]. Research indicates that after treatment, certain
   patients exhibit molecular residual diseases (MRDs) that persist
   undetected through imaging or standard examinations, potentially
   leading to disease relapse or metastasis [[50]5, [51]6]. At the point
   when a solid tumor becomes detectable for recurrence or metastasis via
   conventional methods, the cancer cell count typically exceeds tens of
   millions, rendering eradication highly unlikely, irrespective of the
   available treatment modalities. Consequently, the identification of MRD
   holds substantial clinical significance for early assessment of
   therapeutic response, improving prognostic accuracy or risk
   stratification, and enabling post-treatment surveillance to detect
   potential relapse and guide proactive intervention.

   In general, the detection of MRD depends on the utilization of
   molecular assays that exhibit exceptional sensitivity and specificity,
   such as polymerase chain reaction (PCR)-based and next-generation
   sequencing (NGS)-based technologies. Circulating tumor DNA (ctDNA) is a
   form of circulating cell-free DNA (cfDNA) that tumor cells release
   during apoptosis or necrosis[[52]7–[53]10]. Thus, ctDNA can minor the
   mutation information of tumor DNA in the tumor tissue and can be
   collected noninvasively. Moreover, Its short half-life (< 2 h) makes
   ctDNA ideal for real-time monitoring of tumor changes[[54]11, [55]12].

   Recent advancements in ctDNA examination in peripheral blood highlight
   its potential as a universal biomarker for MRD in solid
   tumors[[56]13–[57]19]. Colorectal cancer is at the forefront of solid
   tumor MRD research. The National Comprehensive Cancer Network (NCCN)
   guidelines for colon cancer now include ctDNA-based MRD assessment,
   using postoperative ctDNA to identify higher recurrence risk in stage
   I-III colon cancer[[58]15]. However, ctDNA use for MRD in ESCC is less
   explored[[59]20–[60]25], with most studies focusing on esophageal
   adenocarcinoma (EAC)[[61]25–[62]34]. However, earlier research has
   indicated that squamous cell carcinomas release more ctDNA compared to
   adenocarcinomas. [[63]25, [64]35–[65]37].

   This study hypothesized that ctDNA-based MRD detection could serve as a
   robust prognostic biomarker for resected ESCC patients. To evaluate
   this, we performed high-depth sequencing on 62 primary tumor tissues,
   108 preoperative and 125 postoperative plasma samples from 125
   surgery-only ESCC patients using targeted NGS panel (437 genes) and
   ultra-high sensitivity Automated Triple Groom Sequencing (ATG-seq)
   panel (196 genes). Our findings establish postoperative
   ctDNA-positivity as both a promising biomarker for MRD detection and an
   effective tool for risk stratification in resected ESCC patients,
   offering clinical insights to future adjuvant therapy decisions
   following curative resection.

Results

Patient characteristics and sample quality control

   In this study, a total of 62 primary tumor tissues, 108 preoperative
   and 125 postoperative plasma samples were collected from 125 ESCC
   patients at pathological stages of II-IVA (the 8th edition of the
   American Joint Committee on Cancer staging system) who underwent
   radical surgery without neoadjuvant and adjuvant therapy, and subjected
   to sequencing. The primary tumor tissue and preoperative plasma served
   as the baseline sample; therefore, all 125 patients had both
   postoperative plasma and a corresponding baseline sample (i.e., either
   primary tumor tissue or preoperative plasma) (Fig. S1a). The median
   follow-up time for the entire cohort was 40.03 months (95% CI:
   38.09–41.66). Patient demographics and baseline clinical
   characteristics are stratified by disease-free survival (DFS) status
   and presented in Table [66]1. All samples used passed quality control
   (Fig. S2), with Q20 scores ≥ 90% and 20X coverage ≥ 90%. The median
   sequencing depths were 266X for blood control, 981X for tumor tissue,
   29,495X for preoperative plasma, and 30,089X for postoperative plasma
   samples.

Table 1.

   Baseline characteristics
   Characteristics No. of With DFS Events n (%) No. of Without DFS Events
   n (%) No. of Total Patients n
   patients 54 71 125
   Median age, years (range) 66(45–78) 63(41–82) 63(41–82)
   Sex
    Female 9(30.00%) 21(70.00%) 30
    Male 45(47.37%) 50(52.63%) 95
   Primary tumor location
    Proximal third 3(60.00%) 2(40.00%) 5
    Middle third 26(42.62%) 35(57.38%) 61
    Distal third 22(43.14%) 29(56.86%) 51
    Gastroesophageal junction 0(00.00%) 2(100.00%) 2
    Multiple primary 3(50.00%) 3(50.00%) 6
   Vascular invasion
    Yes 20(51.28%) 19(48.72%) 39
    No 34(39.53%) 52(60.47%) 86
   Nerve tract invasion
    Yes 27(46.55%) 31(53.45%) 58
    No 27(40.30%) 40(59.70%) 67
   Pathological stage
    II 22(32.35%) 46(67.65%) 68
    III 25(52.08%) 23(47.92%) 48
    IVA 7(77.78%) 2(22.22%) 9
   [67]Open in a new tab

   DFS disease-free survival

Mutation profile in tumor tissue and plasma

   The mutation detection rates for tumor tissue, preoperative plasma, and
   postoperative plasma were 98.39% (61/62), 79.63% (86/108), and 48.00%
   (60/125), respectively (Fig. S1b). Our assay demonstrated a limit of
   detection (LOD) of 0.01% for variant allele frequency (VAF).
   Quantitative mutational features, including total mutation count,
   unique gene involvement, median mutation count (with range), and mean
   VAF values, are detailed in Table S1.

   The mutation spectra of the top 20 high-frequency mutated genes and
   corresponding sunburst diagrams from tumor tissues (Fig. [68]1a),
   preoperative plasma (Fig. [69]1b), and postoperative plasma
   (Fig. [70]1c) samples are provided. High-frequency mutation genes
   detected in the three sample types were consistent with previous
   literature[[71]21, [72]38], such as TP53, CCND1, PIK3CA, CDKN2A, ROS1,
   and ALK etc. TP53 had the highest frequency of mutations in tumor
   tissue, preoperative plasma, and postoperative plasma samples, with
   rates of 90.3%, 56.5%, and 28.0%, respectively. Pathway enrichment
   analysis was performed on high-frequency mutated genes (Fig. S3a-c),
   revealing that the majority of these genes across all three sample
   types were enriched in four key pathways: the cell cycle, cellular
   senescence, the p53 signaling pathway, and the FoxO signaling pathway.
   Mutation genes associated with poor DFS or overall survival (OS) were
   identified in tumor tissues (Table S2), preoperative plasma (Table S3),
   and postoperative plasma (Table S4) samples. Thirteen specific gene
   mutations were significantly associated with inferior survival. The
   overlapping mutated genes across these sample types were visualized
   using an UpSet plot (Fig. S3d). Specifically, TP53 mutations in either
   preoperative or postoperative plasma samples correlated with worse DFS
   (p = 0.004; p = 0.0049) and OS (p = 0.037; p = 0.0086). Additionally,
   SDHA mutations in postoperative plasma were linked to poorer DFS
   (p = 0.02) and OS (p = 0.0049), whereas in preoperative plasma, they
   were only associated with worse OS (p = 0.021). Sunburst plots
   (Fig. [73]1a-c) illustrated the stratified distribution of demographic
   and clinicopathological characteristics and corresponding mean DFS
   across the three sample cohorts. The analysis revealed a male
   predominance in patient distribution, with a balanced representation
   between older (≥ 63 years) and younger (< 63 years) individuals.
   Additionally, the proportion of patients gradually decreased with
   advancing pTNM stage (II to IV), while the mutation-positive rates
   aligned with the aforementioned detection rates.

Fig. 1.

   [74]Fig. 1
   [75]Open in a new tab

   Mutation profiling a-c. Mutation landscapes and hierarchical patient
   data are shown for 62 tumor tissues (a), 108 preoperative (b) and 125
   postoperative (c) plasma samples. Hierarchical patient data are
   presented as sunburst diagrams, with rings representing gender, age,
   pTNM stage, and mutation status from center outward. Color intensity
   reflects the mean DFS duration for each subgroup. DFS, disease-free
   survival

Comparison of genomic DNA from tumor tissue to ctDNA from plasma

   To evaluate the mutation detection concordance between different sample
   types, McNemar’s test for paired samples was first applied to assess
   statistical consistency, followed by percent agreement and
   prevalence-adjusted bias-adjusted kappa (PABAK) to quantify the
   agreement level (Table [76]2). The concordance analysis revealed 91.11%
   agreement (p = 0.62) between tumor tissue and preoperative plasma
   samples in 45 patients (Fig. S1c), with a PABAK of 0.82 indicating
   'almost perfect' agreement according to Landis and Koch criteria
   [[77]39]. Lower agreement rates were observed in other comparisons:
   67.74% (p < 0.001) between tumor tissue and postoperative plasma
   samples (n = 62), and 62.96% (p < 0.001) between preoperative and
   postoperative plasma samples (n = 108).

Table 2.

   Concordance rate of mutation detection
   Preoperative plasma  Tumor tissue        p*       Percent agreement PABAK
                        Negative Positive
    Negative            0        3          0.62     91.11% (41/45)    0.82
    Positive            1        41
   Postoperative plasma Tumor tissue        p*       Percent agreement
                        Negative Positive
    Negative            0        19          < 0.001 67.74% (42/62)
    Positive            1        42
   Postoperative plasma Preoperative plasma p*       Percent agreement
                        Negative Positive
    Negative            21       39          < 0.001 62.96% (68/108)
    Positive            1        47
   [78]Open in a new tab

   ^*McNemar's Chi-squared test with continuity correction, a
   non-significant result (p > 0.05) indicates that both methods have
   equivalent detection rates at the α = 0.05 significance level; Percent
   agreement, calculated as the proportion of concordant pairs among all
   paired samples (n_concordant/n_total); PABAK (prevalence-adjusted
   bias-adjusted kappa), derived as PABAK = 2 × (Percent agreement) − 1,
   where Percent agreement is expressed as a decimal

   There were two distinct sources of mutations in ctDNA present in
   plasma: originating from the tracked mutations derived from tumor
   tissue, i.e., tumor-informed mutations; and unique variations solely
   detected in plasma, i.e., tumor-naive mutations. In order to clarify
   the sources of mutations in ctDNA, we compared the mutations detected
   in ctDNA-positivity plasma samples and their corresponding paired tumor
   tissues. Among the preoperative ctDNA-positivity plasma samples, 42 had
   paired tumor tissue samples that were sequenced (Fig. S1c,
   Fig. [79]2a-b). In the 42 paired samples, 254 mutations involving 131
   genes were identified in tumor tissue samples, and 104 mutations
   involving 41 genes in paired preoperative plasma samples. And for the
   42 preoperative plasma samples, a total of 87 tumor-informed mutations
   were detected, while 17 tumor-naive mutations were found. Among the
   postoperative ctDNA-positivity plasma samples, 43 had paired tumor
   tissue samples that were sequenced (Fig. S1d, Fig. [80]2c-d). In the 43
   pairs of samples, 273 mutations involving 137 genes were identified in
   tumor tissue samples, and 86 mutations involving 39 genes in
   postoperative plasma samples. And for the 43 postoperative plasma
   samples, 74 tumor-informed mutations were detected, while 12
   tumor-naive mutations were found. Among the 45 patients with matched
   tumor tissue, preoperative plasma, and postoperative plasma sequencing
   data, the two mutation sources across all three sample types using Venn
   diagrams were characterized (Fig. S4a).

Fig. 2.

   [81]Fig. 2
   [82]Open in a new tab

   Comparison of genomic DNA from tumor tissue and ctDNA from plasma a-b.
   The number of mutations discovered in 42 paired tumor tissue samples
   and preoperative plasma samples is presented by a Venn diagram (a) or
   by cases (b). c-d. The number of mutations discovered in 43 paired
   tumor tissue samples and postoperative plasma samples is presented by a
   Venn diagram (c) or by cases (d). e–f. Kaplan–Meier survival curves for
   DFS (e), and Kaplan–Meier survival curves for OS (f) among patients
   with tumor-informed, tumor-naïve, and negative ctDNA mutations. DFS,
   disease-free survival; OS, overall survival

   Next, we explored survival differences among patients with
   tumor-informed, tumor-naïve, and negative ctDNA mutations. In the 62
   patients with paired tumor tissue samples and postoperative plasma
   samples sequenced (Fig. S1d), 19 were ctDNA-negative, while 43 were
   ctDNA-positive. Of the ctDNA-positive patients (mutation details in
   Fig. [83]2d), 10 harbored tumor-naïve mutations (tumor-naïve group),
   whereas the remaining 33 exhibited only tumor-informed mutations
   (tumor-informed group). Patients in the tumor-informed (DFS: hazard
   ratio [HR] 5.59, 95% CI 2.66–11.73, p < 0.001; OS: HR 6.96, 95% CI
   3.13–15.49, p < 0.001) and tumor-naïve (DFS: HR 5.34, 95% CI
   1.36–21.03, p = 0.0025; OS: HR 5.89, 95% CI 1.31–26.42, p = 0.0041)
   groups had poorer DFS and OS than those in the ctDNA-negative group.
   However, no significant difference was observed in DFS (p = 0.98) and
   OS (p = 0.89) between patients in the tumor-informed and tumor-naïve
   groups (Fig. [84]2e-f). In consideration of the same prognosis of
   tumor-informed and tumor-naïve ctDNA mutation, we defined
   ctDNA-positivity as either a tumor-informed or a tumor-naïve ctDNA
   mutation detected.

Correlation between ctDNA detection and pathological stage

   An increase in preoperative ctDNA-positivity rates was observed with
   increasing pathological stage, with rates of 70.97% (44/62), 92.11%
   (35/38), and 87.50% (7/8) for pathological stage II, III, and IVA
   patients, respectively (p = 0.027) (Fig. [85]3a). Numerical but not
   significant differences in the relative abundance of
   mutations (quantified as maxVAF × cfDNA concentration) and the number
   of mutations of ctDNA from preoperative plasma were identified in
   patients with different pathological stages (Fig. [86]3b-c). In terms
   of postoperative plasma, there was an increase in the detection rate of
   ctDNA mutations as overall pathological stage increased. Rates were
   41.18% (28/68), 52.08% (25/48), and 77.78% (7/9) for patients with
   pathological stages II, III, and IVA, respectively (p = 0.10)
   (Fig. [87]3d). Similarly, as stage advanced, both the relative
   abundance (p = 0.043) and number of mutations (p = 0.047) in ctDNA from
   postoperative plasma increased (Fig. [88]3e-f).

Fig. 3.

   [89]Fig. 3
   [90]Open in a new tab

   Correlation between ctDNA detection and pathological stage a-c. The
   mutation detection rate (a), relative abundance (b) and number of
   mutations of ctDNA (c) from preoperative plasma in patients with
   different pathological stages were compared. d-f. The mutation
   detection rate (d), relative abundance (e) and number of mutations of
   ctDNA (f) from postoperative plasma in patients with different
   pathological stages were compared. g. The mutation clearance rate in
   patients with different pathological stages was compared. ctDNA-neg,
   ctDNA-negativity; ctDNA-pos, ctDNA-positivity. Fig. b and e are Tukey
   box plots. Fig. c and f are bar plots with error bars (mean ± 95% CI)

   We also evaluated the conversion patterns of ctDNA detection status in
   108 patients with matched pre- and postoperative plasma samples (Fig.
   S1e). Among 22 preoperative ctDNA-negative patients, 21 (95.5%) showed
   persistent negativity, whereas 1 (4.5%) developed ctDNA-positivity
   postoperatively. The single ctDNA-positive patient demonstrated worse
   median DFS (7.76 vs not reached [NR] months, p = 0.0023) and OS (21.4
   vs NR months, p = 0.0031) compared to persistent negativity patients.
   Of 86 preoperatively ctDNA-positive patients, 39 (45.4%) converted to
   negative status (ctDNA-clearance) and 47 (54.6%) remained positive
   (ctDNA-nonclearance). The ctDNA clearance rate
   (ctDNA-clearance/[ctDNA-clearance + ctDNA-nonclearance]) showed a
   decreasing trend with advancing pathological stage: 47.73% (21/44) in
   stage II, 48.57% (17/35) in stage III, and 14.29% (1/7) in stage IVA
   patients (p = 0.28) (Fig. [91]3g).

Correlation between ctDNA detection status and risk of recurrence or death

   The recurrence rate of the total cohort was 43.20% (54/125) and an
   increase in recurrence was observed with pathological stage
   progression, with rates of 32.35% (22/68), 52.08% (25/48), and 77.78%
   (7/9) for pathological stage II, III, and IVA patients, respectively
   (Fisher’s Exact test, p = 0.010). For outcomes based on preoperative
   plasma, patients with preoperative ctDNA-positivity had a higher
   recurrence rate than those with preoperative ctDNA-negativity (50.0%
   (43/86) vs. Negativity: 18.18% (4/22), p = 0.014) (Fig. [92]4a). Also,
   survival analyses revealed a statistically poorer DFS (p = 0.016)
   (Fig. [93]4b) and a trend toward worse OS (p = 0.071) (Fig. [94]4c) in
   patients with preoperative ctDNA-positivity.

Fig. 4.

   [95]Fig. 4
   [96]Open in a new tab

   Correlation between ctDNA detection status and risk of recurrence or
   death a-c. Overall relapse proportion (a), Kaplan–Meier survival curves
   for DFS (b), and Kaplan–Meier survival curves for OS (c) between
   patients with ctDNA-negativity and ctDNA-positivity preoperatively.
   d-f. Overall relapse proportion (d), Kaplan–Meier survival curves for
   DFS (e), and Kaplan–Meier survival curves for OS (f) between patients
   with ctDNA-negativity and ctDNA-positivity postoperatively. g-i.
   Overall relapse proportion (g), Kaplan–Meier survival curves for DFS
   (h), and Kaplan–Meier survival curves for OS (i) between patients with
   ctDNA-clearance and patients with ctDNA-nonclearance. HR, hazard ratio.
   DFS, disease-free survival; OS, overall survival; ctDNA-neg,
   ctDNA-negativity; ctDNA-pos, ctDNA-positivity

   Notably, patients with postoperative ctDNA-positivity had a higher
   recurrence rate than those with ctDNA-negativity (66.67% (40/60) vs.
   21.54% (14/65), p < 0.001) (Fig. [97]4d). Moreover, DFS was
   significantly worse in patients with postoperative ctDNA-positivity
   (DFS: 20.30 vs. NR months, HR: 4.58, 95% CI: 2.65–7.92, p < 0.001)
   (Fig. [98]4e), as was their OS (mOS: 38.22 vs. NR months, HR: 5.39, 95%
   CI: 2.96–9.80, p < 0.001) (Fig. [99]4f). A higher recurrence risk
   (Positivity: 68.09% (32/47) vs. Negativity: 28.21% (11/39), p < 0.001)
   (Fig. [100]4g), as well as worse DFS (mDFS: 20.44 vs. NR months, HR:
   3.55, 95% CI: 1.95–6.47, p < 0.001) (Fig. [101]4h) and OS (mOS: 35.39
   vs. NR months, HR: 4.85, 95% CI: 2.49–9.43, p < 0.001) (Fig. [102]4i)
   were also observed in patients without ctDNA-clearance after surgery.

   The results of multivariate Cox regression analysis reinforced the
   potential of postoperative ctDNA-positivity as a reliable indicator for
   MRD, as it was identified as an independent prognostic factor for DFS
   (HR: 4.10, 95% CI: 2.03–8.29, p < 0.001) (Table [103]3) and OS (HR:
   5.38, 95% CI: 2.65–10.95, p < 0.001) (Table [104]3). The prognostic
   value of postoperative ctDNA-positivity was also explored in different
   subgroups (Fig. [105]5a-f). Compared with postoperative ctDNA-negative
   patients, pathological stage II and III patients with postoperative
   ctDNA-positivity all had a worse prognosis, including lower DFS (stage
   II: p = 0.0079; stage III: p < 0.001) (Fig. [106]5a-b) and OS (stage
   II: p = 0.015; stage III: p < 0.001) (Fig. [107]5d-e). There were a few
   patients with stage IVA disease, and no significant differences in DFS
   (p = 0.68) and OS (p = 0.68) were detected for stage IVA patients
   (Fig. [108]5c, f).

Table 3.

   DFS and OS analysis of clinicopathologic variables and ctDNA status in
   ESCC
   Variables No. of patients Univariate analysis Multivariate analysis
   HR (95% CI) p HR (95% CI) p
   DFS
    Sex (male vs. female) 95/125 1.85(0.91–3.79) 0.091
    Age (≥ 63 vs. < 63) 68/125 1.19(0.69–2.04) 0.53
    Stage III-IVA vs. Stage II 57/125 2.17(1.26–3.75) 0.0052
   1.67(0.93–2.99) 0.084
    Nerve tract invasion 58/125 1.20(0.71–2.05) 0.50
    Vascular invasion 39/125 1.44(0.83–2.51) 0.19
    Preoperative ctDNA-positivity 86/108 3.28(1.18–9.13) 0.023
   1.28(0.41–4.06) 0.67
    Postoperative ctDNA-positivity 60/125 4.72(2.55–8.71)  < 0.001
   4.10(2.03–8.29)  < 0.001
   OS
    Sex (male vs. female) 95/125 2.03(0.90–4.54) 0.086
    Age (≥ 63 vs. < 63) 68/125 1.41(0.77–2.56) 0.26
    Stage III-IVA vs. Stage II 57/125 1.87(1.03–3.37) 0.039
   1.69(0.94–3.07) 0.082
    Nerve tract invasion 58/125 1.28(0.71–2.29) 0.41
    Vascular invasion 39/125 1.42(0.78–2.59) 0.26
    Preoperative ctDNA-positivity 86/108 2.51(0.89–7.06) 0.082
    Postoperative ctDNA-positivity 60/125 5.55(2.73–11.26)  < 0.001
   5.38(2.65–10.95)  < 0.001
   [109]Open in a new tab

   DFS disease-free survival, OS overall survival, ctDNA circulating tumor
   DNA, ESCC esophageal squamous cell carcinoma

Fig. 5.

   [110]Fig. 5
   [111]Open in a new tab

   Subgroup analysis and TNMB staging system a-c. Kaplan–Meier survival
   curves for DFS in patients with pathological stage II (a), stage III
   (b), and stage IVA (c). d-f. Kaplan–Meier survival curves for OS in
   patients with pathological stage II (d), stage III (e), and stage IVA
   (f). g-h. Exploratory Kaplan–Meier analysis for DFS (g) and OS (h)
   among patients in group II-neg, group II-pos, group III-neg, group
   III-pos, group IVA-neg, and group IVA-pos. i-j. The integration of
   Kaplan–Meier analysis for DFS (i) and OS (j) in patients divided by TNM
   staging and the TNMB staging system. k-l (k) DFS and (l) OS
   time-dependent ROC comparing TNM (blue line) vs. TNMB (green line)
   staging system. neg, ctDNA-negativity; pos, ctDNA-positivity; TNM,
   Tumor-Node-Metastasis; TNMB, Tumor-Node-Metastasis-Blood; DFS,
   disease-free survival; OS, overall survival; ROC, receiver operating
   characteristic; AUC, area under the receiver operating characteristic
   curve

   The efficacy of postoperative ctDNA-positivity for MRD detection was
   further evaluated. The performance of the postoperative ctDNA assay for
   recurrence prediction within one year of follow-up was 83.33%
   sensitivity and 60.40% specificity, with an overall accuracy of 64.8%
   (81/125) (Table S5). As the follow-up time increased, the sensitivity
   and negative predictive value (NPV) of postoperative ctDNA positivity
   for recurrence prediction decreased, but the specificity and positive
   predictive value (PPV) increased. Relatively high values of sensitivity
   and NPV for the postoperative ctDNA assay were observed for recurrence
   prediction within one year, two years, or three years of follow-up.

Combination of the TNM stage system and ctDNA detection status

   Given the above findings, we combined the postoperative ctDNA detection
   status with the traditional pathological tumor-node-metastasis (TNM)
   staging to assess patient prognosis, which was defined as
   tumor-node-metastasis-blood (TNMB) staging. Based on exploratory
   results (Fig. [112]5g-h), the specific staging criteria for TNMB were
   as follows: pathological stage II as well as ctDNA-negativity and
   pathological stage III as well as ctDNA-negativity were defined as TNMB
   stage 1; pathological stage II as well as ctDNA-positivity and
   pathological stage IVA as well as ctDNA-negativity were defined as TNMB
   stage 2; and pathological stage III as well as ctDNA-positivity and
   pathological stage IVA as well as ctDNA-positivity were defined as TNMB
   stage 3. We found that this new TNMB staging system was more effective
   in distinguishing patient prognosis than the traditional TNM staging
   (Fisher's Exact Test, p < 0.001) (Fig. [113]5i-j). Time-dependent
   receiver operating characteristic (ROC) curves demonstrated the
   discriminative ability of TNMB and TNM staging for DFS and OS at
   different time points (Fig. [114]5k-l). TNMB demonstrated consistently
   higher predictive accuracy than TNM staging, achieving higher C-indices
   (DFS: 0.80 vs. 0.65; OS: 0.77 vs. 0.69). These findings were further
   validated in an external cohort[[115]40] of ESCC patients receiving
   definitive chemoradiotherapy combined with toripalimab (Fig. S4b-c),
   where TNMB maintained superior prognostic discrimination (PFS: 0.75 vs.
   0.71; OS: 0.80 vs. 0.72). The results suggest that the TNMB staging
   system can enhance and refine traditional TNM staging in predicting
   patient prognosis.

Discussion

   Our study analyzed gene mutations in 62 primary tumor tissues, 108
   preoperative plasma samples, and 125 postoperative samples from 125
   ESCC patients who underwent surgery. Using ATG-seq (196 genes) with
   50 ng of ctDNA per sample, we achieved a sequencing depth
   of ~ 30,000 × and a VAF detection Limit of 0.01%. Combining
   tumor-informed and tumor-naive methods, we found high concordance
   between tumor tissue and preoperative plasma, as well as high mutation
   detection rates in both preoperative and postoperative plasma.
   Preoperative and postoperative ctDNA-positivity, along with
   ctDNA-nonclearance, were linked to higher recurrence rates and poorer
   DFS and OS. Postoperative ctDNA-positivity was identified as an
   independent prognostic indicator for DFS and OS, acting as a
   dependable biomarker for identifying MRD in ESCC patients.

   Our study found postoperative ctDNA testing achieved 83.33% sensitivity
   and 60.40% specificity for one-year recurrence prediction (64.8%
   accuracy). Compared to imaging modalities[[116]41, [117]42] which
   showed lower sensitivity (N-stage: 35–42%; M-stage: 46–69%) but higher
   specificity (N-stage: 87–93%; M-stage: 74–93%) with 63–84% accuracy,
   ctDNA provided superior sensitivity for MRD detection while maintaining
   acceptable accuracy despite lower specificity. The results further
   substantiated the utility of postoperative ctDNA testing for MRD
   detection and its potential to enhance postoperative surveillance
   protocols. We propose integrating longitudinal plasma ctDNA monitoring
   into follow-up protocols, where positive results could trigger
   confirmatory imaging. This approach may enable earlier recurrence
   detection while reducing unnecessary resource utilization, though
   further validation is required. Regarding ctDNA-guided adjuvant
   therapy, our comprehensive literature review reveals that there are no
   large-scale randomized controlled trials demonstrating overall survival
   benefits from adjuvant radiotherapy or chemotherapy in ESCC, and these
   treatments remain non-standard[[118]43, [119]44]. However, the landmark
   CheckMate 577 trial[[120]45] revealed that adjuvant nivolumab
   administered for one year significantly improved DFS in patients with
   locally advanced esophageal or gastroesophageal junction cancer who
   achieved R0 resection but had residual disease (non-pathologic complete
   response) after neoadjuvant chemoradiotherapy. In a phase II clinical
   trial[[121]46] comparing neoadjuvant PD-L1 inhibitor (socazolimab) plus
   chemotherapy versus chemotherapy alone in locally advanced ESCC, Yue et
   al.[[122]47] subsequently performed MRD detection through postoperative
   cfDNA analysis. Their results showed that among 11 MRD-positive
   patients, none of the 7 patients who received adjuvant socazolimab
   experienced disease progression, whereas 3 of the 4 untreated patients
   developed recurrence during the median follow-up period of 17 months
   (HR: 0.0032; 95% CI, 0.003–0.390; p = 0.007). Moving forward, clinical
   investigations should evaluate the potential role of postoperative
   ctDNA status in stratifying patients for adjuvant immunotherapy
   regimens.

   The TNMB staging system extends the traditional TNM framework by
   incorporating tumor biological behavior and molecular characteristics
   (B), in addition to primary tumor extent (T), nodal involvement (N),
   and distant metastasis (M). Building on landmark MRD status assessed
   via PROPHET, Chen et al.[[123]48] first proposed the TNMB staging
   system for non-small cell lung cancer and found that TNMB staging
   showed a better prognostic effect than traditional TNM staging. While
   the TNMB criteria in our study differ from theirs due to variations in
   cancer type, our results similarly confirmed that the TNMB system
   provides significantly better prognostic stratification for ESCC
   patients (p < 0.001). Notably, this new system showed significantly
   higher predictive accuracy compared to the conventional TNM staging
   (DFS: 0.80 vs. 0.65; OS: 0.77 vs. 0.69). This improvement likely stems
   from the ability of postoperative ctDNA-positivity to identify patients
   with MRD who are resistant to surgical intervention. The underlying
   rationale parallels that of ypTNM staging, which identifies
   treatment-resistant patients with residual tumors after neoadjuvant
   therapy. The ypTNM staging paradigm[[124]49] exhibits globally
   depressed survival outcomes, with early-stage patients significantly
   worse than pTNM counterparts while advanced-stage patients show
   comparable or inferior survival to conventional staging. Enhanced
   discrimination between prognostic groups enables more precise
   stratification for adjuvant therapy decision-making, thereby advancing
   the goals of personalized precision medicine. A review of the existing
   literature reveals that the TNMB stage system for esophageal cancer has
   been rarely explored, and its introduction might open up new directions
   and perspectives for future esophageal cancer research and treatment.

   Existing literature [[125]20–[126]25] on ctDNA detection in ESCC
   patients has utilized fixed panels, with various factors significantly
   influencing the probability of detecting ctDNA. These elements
   encompass the ctDNA concentrations in a plasma sample, the sequencing
   depth at particular mutant sites, and the scope of analysis, which
   pertains to the number of mutations monitored[[127]50]. Previous
   studies [[128]21, [129]22, [130]37] have reported mutation detection
   rates of 53.3% ~ 71.7% for preoperative plasma, and 14.2% ~ 15.8% for
   postoperative plasma in ESCC patients. In our study, the detection rate
   of ctDNA was relatively high at 79.63% (86/108) in the total population
   preoperatively and 48.00% (60/125) postoperatively, which might be
   attributed to several key factors. Firstly, our study achieved a highly
   sensitive LOD for VAF of 0.01% in ctDNA testing, allowing for the
   identification of even trace amounts of ctDNA. However, previous
   studies regarding NGS MRD detection in esophageal cancer have rarely
   reported LOD or had a LOD of 0.1% ~ 0.2%[[131]23, [132]24, [133]27].
   Second, our deep sequencing approach (a mean depth of ~ 30,000 ×)
   provided thorough and in-depth analysis of ctDNA mutations, enhancing
   our ability to detect and characterize genetic alterations accurately.
   Nevertheless, previous studies of NGS MRD detection in esophageal
   cancer have rarely reported sequencing depth or had a relatively low
   mean sequencing depth of ~ 7000 × [[134]31]. Additionally, utilization
   of a large NGS panel (437 genes) expanded the scope of genomic
   alterations that could be captured and analyzed. Last, integration of
   tumor-informed and tumor-naive methods played a significant role in
   optimizing ctDNA detection and further improving the overall detection
   rate in our research. In general, researchers[[135]20–[136]25] employ
   either tumor-informed or tumor-naive approaches for analyzing MRD. The
   tumor-informed approach utilizes mutation data from primary tumor
   tissue, forming a personalized mutation profile for each patient and
   using it for MRD monitoring. The tumor-naive approach aims to achieve
   the highest sensitivity by using UMIs, increasing sequencing depth, and
   employing advanced algorithms, without relying on any information from
   tumor tissue. This approach solely relies on blood detection to obtain
   accurate results. Our analysis revealed that mutations identified by
   both methods correlated with worse DFS and OS, yet relying solely on
   tumor-informed analysis would have missed these high-risk patients. And
   integration of these methods can improve the detection of ctDNA
   mutations, providing valuable insights for MRD detection.

   Notably, similar observations have been reported in colorectal cancer.
   Martínez-Castedo et al.[[137]51] showed that both tumor-informed and
   tumor-naïve approaches are clinically valuable for assessing
   postoperative risk and making decisions about adjuvant chemotherapy,
   emphasizing that although tumor-informed methods currently provide
   greater accuracy in MRD detection, tumor-naive techniques are becoming
   more popular because of their ease of use and enhanced performance. In
   our study, we enhanced the tumor-naïve methodology by integrating UMIs,
   increasing sequencing depth, and employing advanced bioinformatics
   algorithms, thereby achieving a detection sensitivity of 0.01% VAF. Our
   hybrid strategy represents a practical compromise, helping to identify
   clinically relevant mutations without requiring tumor tissue
   sequencing, while still maintaining reliable detection performance.

   In this study, we designed NGS gene panels as pan-cancer detection
   tools, encompassing full exonic regions and critical intronic regions
   of cancer-related genes. The panel includes not only well-characterized
   esophageal cancer-associated genes[[138]52] (e.g., TP53, CDKN2A,
   PIK3CA, NOTCH1, KMT2D) but also other clinically relevant genes across
   multiple tumor types, ensuring broad coverage for Liquid biopsy
   applications. Our analysis revealed 13 specific gene mutations
   significantly correlated with adverse survival outcomes: TP53, SDHA,
   ATR, NFKBIA, FANCI, ATRX, LRP1B, NFE2L2, CDKN2A, TUBB3, ATM, PIK3CA,
   VHL, and ATR. Among these, TP53 and SDHA mutations in plasma showed
   comparatively significant prognostic value. Supporting this finding, Li
   et al.[[139]53] reported that the TP53 hotspot mutation p.R213* was
   found to independently predict lower survival rates, with a hazard
   ratio of 3.37, even after adjustment for clinical variables.
   Conversely, the role of SDHA mutations in ESCC remains understudied in
   the literature. Based on these preliminary findings, we suggest
   considering the inclusion of these candidate genes in future ESCC
   prognostic panels to enhance prognostic accuracy and clinical utility,
   pending further validation in larger cohorts.

   Most studies [[140]20, [141]21, [142]23–[143]25] have seldom shown
   poorer prognoses in patients with preoperative ctDNA-positivity or
   persistent ctDNA levels. For instance, Iwaya[[144]20] et al. enrolled
   36 ESCC patients and discovered no notable difference in OS between
   those with ctDNA-positive and ctDNA-negative pretreatment plasma
   (p = 0.46). In our current study, there were 108 patients with
   preoperative plasma available, and we found that those with
   preoperative ctDNA-positivity had a higher pathological stage
   (p = 0.027) and recurrence rate (p = 0.014), worse DFS (p = 0.016) and
   a trend toward poorer OS (p = 0.071). Moreover, compared with
   preoperative plasma, a notable decrease in the mutation detection rate
   of ctDNA from postoperative plasma was observed (preoperative: 79.63%
   vs. postoperative: 48.00%, Chi-squared test, p < 0.001), probably
   caused by the surgery and reduction in tumor burden. Patients with
   ctDNA-nonclearance also had a higher recurrence rate (p < 0.001), and
   poorer DFS (p < 0.001) and OS (p < 0.001). To date, there is a scarcity
   of literature reporting on the clinical significance of
   ctDNA-nonclearance in patients with ESCC. These results highlight the
   importance of ctDNA in evaluating tumor load and its potential role as
   a prognostic marker for informing clinical treatment choices for
   patients.

   In our study, tumor tissue sequencing was not performed for all
   patients due to practical constraints. Instead, we utilized a
   combination of tumor tissue (n = 62) and preoperative plasma (n = 108)
   sequencing as baseline samples, while postoperative plasma (n = 125)
   sequencing was used to assess MRD. All 125 patients with postoperative
   plasma data had matched baseline sequencing from either tumor tissue or
   preoperative plasma. Specifically, among these patients, 45 had both
   tumor tissue and preoperative plasma sequencing, 17 had only tumor
   tissue sequencing, and 63 had only preoperative plasma sequencing. This
   design enabled multiple paired analyses: the 45 patients with both
   sample types allowed us to evaluate mutation concordance between tissue
   and preoperative plasma; the 62 patients with tumor tissue data (45
   plus 17) supported survival comparisons between tumor-informed,
   tumor-naïve, and ctDNA-negative subgroups; and the 108 patients with
   pre- and postoperative plasma samples (45 plus 63) facilitated ctDNA
   clearance analysis. Although universal tumor sequencing could have
   increased the statistical power of tissue-based analysis, the primary
   findings regarding postoperative ctDNA-based MRD detection remain
   robust, given the high concordance between tumor tissue and
   preoperative plasma (91.11%) and the comparable survival outcomes in
   patients with tumor-naïve and tumor-informed mutations. The relatively
   low concordance rates between preoperative and postoperative plasma
   (62.96%) and between tumor tissue and postoperative plasma (67.74%) are
   also noteworthy. These discrepancies may be attributed to reduced tumor
   burden following surgery, tumor heterogeneity, or tumor evolution.
   Supporting this observation, in the 45 patients with matched tumor
   tissue, preoperative, and postoperative samples (Fig. S4a),
   postoperative plasma acquired 12 additional mutations not detected in
   tumor tissue and 22 new mutations absent in preoperative plasma.

   There are several constraints in our research. Our study was conducted
   retrospectively, which may introduce bias and limit the strength of our
   conclusions. Future prospective studies will provide more robust
   evidence. Additionally, our study was limited to ctDNA testing at a
   single time point after surgery, without multiple time point sampling.
   Consequently, we could not assess whether ctDNA can detect MRD earlier
   than imaging examinations. Lastly, we limited our study to patients who
   did not receive neoadjuvant or adjuvant therapy (between September 12,
   2016, and January 8, 2020), following the treatment guidelines of that
   time period. As a consequence, we were unable to evaluate the
   recurrence predictive value of ctDNA for patients with resectable ESCC
   who received neoadjuvant or adjuvant therapy.

   In conclusion, we found that preoperative ctDNA-positivity,
   postoperative ctDNA-positivity and ctDNA-nonclearance were associated
   with worse DFS and OS and that postoperative ctDNA was an independent
   prognostic factor for DFS and OS in ESCC patients after esophagectomy.
   The TNMB staging system demonstrates greater efficacy in distinguishing
   patient prognosis than traditional TNM staging. Postoperative
   ctDNA-positivity can be a potent prognostic biomarker for MRD detection
   in ESCC.

Methods

Study design

   Conducted at Sun Yat-sen University Cancer Center, this observational
   cohort study included 125 patients with resectable ESCC between
   September 12, 2016, and January 8, 2020. Eligible patients had
   resectable ESCC with a pathological stage of II-IVA according to the
   8th edition of the American Joint Committee on Cancer staging system.
   All patients underwent radical surgery, consisting of either McKeown or
   Ivor Lewis esophagectomy, which included a two-field lymphadenectomy
   with total mediastinal lymph node dissection, and did not receive any
   neoadjuvant or adjuvant therapy. Formalin-Fixed Paraffin-Embedded
   (FFPE) tumor tissues and plasma samples (before and 7–14 days
   post-surgery) were obtained from the Sun Yat-sen University Cancer
   Center Bio-bank for DNA extraction. Follow-ups were every three months
   in the first year, then biannually until death. The study received
   ethics approval by the Ethics Committee of Sun Yat-sen University
   Cancer Center and exemption from informed consent was granted (Consent
   Letter Number: B2022-090–01; Consent Date: Jan 21, 2022). The
   registration number for our study on ClinicalTrials.gov is
   ChiCTR2400093167.

Next-gen sequencing applied to genomic DNA and ctDNA

   DNA was extracted from FFPE sections and whole blood control samples
   using the QIAamp DNA FFPE Tissue kit (Qiagen) and the DNeasy Blood and
   Tissue kit (Qiagen), respectively. The QIAamp Circulating Nucleic Acid
   Kit (Qiagen) was used to extract plasma ctDNA from every plasma sample.
   A 437-gene targeted NGS panel from Nanjing Geneseeq Technology Inc.
   (Table S6) was employed for analyzing tumors and white blood cells,
   whereas a 196-gene panel (Table S7) was used for plasma ctDNA. The KAPA
   Hyper Prep kit from KAPA Biosystems was used to prepare sequencing
   libraries. A customized library preparation using a bi-barcoding system
   and an ultradeep sequencing method known as ATG-Seq (Nanjing Geneseeq
   Technology Inc.) was utilized for the ctDNA samples. To reduce errors
   from PCR, hybridization, damage, sequencing, and contamination, and to
   prevent mutations from non-tumor sources in ctDNA, we implemented the
   following procedures: (i) sequencing the ctDNA fragment at an
   approximate depth of 30,000 ×, resulting in redundant DNA molecules;
   (ii) utilizing mapping positions and a bi-barcode system to enhance the
   representation of unique DNA molecules; and (iii) employing a
   duplex-assisted decoder system designed to filter mapping and
   sequencing artifacts. Sequencing was carried out on the Illumina
   HiSeq4000 platform, with subsequent data analysis done as previously
   detailed [[145]54]. For plasma samples containing up to 50 ng of ctDNA,
   the LOD for VAF was 0.01%, determined using approximately 30,000 × deep
   sequencing of DNA mixtures from two reference human DNA samples
   (NA19240 and NA18535).

Data processing

   Trimmomatic [[146]55] was employed for quality control of FASTQ files,
   removing leading or trailing bases with low quality (below a score of
   20) or N bases. The paired-end reads were subsequently aligned to the
   human reference genome (build hg19) using the Burrows-Wheeler Aligner
   (BWA) with specified parameters [[147]56]. Picard was used for PCR
   deduplication, and GATK3 was utilized for local realignment around
   indels and recalibration of base quality scores. VarScan2 [[148]57]
   with default settings was used to identify single-nucleotide variations
   (SNVs) and insertions/deletions in tumor tissue samples. FACTERA
   [[149]58] was used with default settings to identify genomic fusions.

   For ctDNA samples, single-stranded consensus sequences (SSCSs) were
   created by gathering all read pairs with identical mapping positions
   and organizing them into distinct SSCS families sharing the same UMI
   barcode sequences at both ends. At least two reads were required to
   support a consensus read. Once the SSCS sequence was constructed, two
   SSCS read pairs with swapped UMI barcode sequences and identical
   mapping positions were combined into a single duplex consensus sequence
   (DCS). Subsequently, a local bioinformatics polishing pipeline was
   employed to detect somatic variants in ctDNA after excluding germline
   variants using normal control DNA. Mutations found in the corresponding
   tumor DNA, supported by at least one unique consensus mutant allele
   read and meeting the polishing criteria, were considered present.
   Tumor-informed ctDNA positivity was determined by detecting one or more
   mutations found in the corresponding tumor samples. ctDNA variants
   absent in the primary tumor are termed tumor-naive ctDNA positivity if
   the following strict criteria are satisfied, (i) VAF is at least 0.01%,
   with supporting reads of 5 or more, and a depth of at least 100x; (ii)
   absent from our earlier published database of clonal hematopoiesis
   variants [[150]59]; (iii) VAF is 0% in the matched peripheral blood
   leukocyte sample.

Statistical analyses

   DFS was calculated from the surgery date to the first confirmed
   radiographic recurrence, death, or the most recent follow-up. OS was
   measured from the surgery date to either the date of death or the last
   follow-up. Categorical variables were analyzed using the Chi-square
   test and Fisher’s Exact test. The t-test was used to compare normally
   distributed continuous variables, whereas the rank-sum test was applied
   to those that were not normally distributed. DFS and OS were estimated
   using the Kaplan–Meier method, and the log-rank test was used to
   compare the resulting curves. The influence of different factors on
   survival was evaluated through univariate and multivariate analyses
   using the Cox proportional hazards model. An analysis of subgroups was
   performed to evaluate the prognostic significance of postoperative
   ctDNA positivity in patients categorized by pathologic TNM stages
   (stage II, stage III, and stage IVA). Clinical variables that were
   statistically significant in the univariate analysis were used for
   multivariate analysis. We transformed the TNMB and TNM staging systems
   into dummy variables and then compared their prognostic performance
   using time-dependent ROC curves and Harrell's C-index. All p-values
   were calculated using two-tailed tests, with significance determined at
   p < 0.05. The statistical analysis was conducted using R software
   (version 4.2.3).

Supplementary Information

   [151]43556_2025_310_MOESM1_ESM.docx^ (2.8MB, docx)

   Supplementary Material 1. Table S1: Quantitative mutation features
   across sample types Table S2: Influence of gene mutations from tumor
   tissue samples on DFS and OS Table S3: Influence of gene mutations from
   preoperative plasma samples on DFS and OS Table S4: Influence of gene
   mutations from postoperative plasma samples on DFS and OS Table S5:
   Performance of the postoperative ctDNA assay for recurrence Table S6:
   Gene list of the Geneseeq Prime™ 437-gene panel Table S7: Gene list of
   196-gene panel Figure S1: Study cohort and sample analysis overview
   Figure S2: Quality control metrics for sequencing data Figure S3:
   Enriched pathway and survival-associated gene overlap. Figure S4:
   Mutation comparison across sample types and external validation of TNMB
   staging system.

Acknowledgements