2.1 Clinicopathological characteristics of patients enrolled with resectable GC

A total of 59 patients were enrolled in this study, with 168 longitude blood samples and 53 corresponding primary tumor samples. Fifty-six patients had baseline blood samples collected pre-operatively and 41 patients had blood samples collected in 1-month post-operatively during follow-up (Figure S1A,B). Information on demographics and clinicopathological characteristics of patients are presented in Table 1. Most tumors were located in the stomach rather than the gastroesophageal junction. Approximately 40.68% of tumors had a maximum diameter of over 5 cm. Pathological examination revealed poorly differentiated tumors in 86.44% of cases, while the remaining cases were moderately to highly differentiated. Regarding the Lauren's classification, 15 patients (25.42%) had a diffuse type and 16 patients (27.12%) had intestinal type. Most patients were detected with lymphatic invasion (LVI) (71.19%). More than half (55.93%) of patients were diagnosed with an advanced stage (pathological tumor-node-metastasis [pTNM] III-IV). Based on the The Cancer Genome Atlas Program (TCGA) molecular classification, two tumors were classified as Epstein–Barr virus (EBV) type (3.39%) and four tumors were classified as microsatellite instability (MSI) type (6.78%). The chromosomal instability (CIN) type accounted for 59.32% of total cases, while the remaining cases were classified as genomic stability (GS) type (30.51%). All 59 patients received surgery with curative intent, and 61.02% of them received adjuvant chemotherapy. As of the data cut-off on February 18, 2022, the median follow-up time was 746 (24–969) days. Twenty-five (42.37%) patients had a recurrence, and 18 (30.51%) patients died from tumor-related causes.

2.2 The mutational landscape of genomic alterations in ctDNA and tumor tissue of resectable GC

To assess the concordance of genomic alterations between ctDNA in plasma samples from different phases and tumor tissue, genes were sequenced via the Roche AVENIO surveillance panel, a pan-cancer panel that spans 198 kb and encompasses 197 genes. According to the captured NGS analysis, there was only a small subset of patients with detectable copy number variation in both blood samples (3.4%) and tumor tissue (7.5%) (Figure 1A). Notably, there was a distinct difference in Top20 genes ranked by mutation frequency between pre-operative and post-operative blood samples (Figure 1B–D). Nevertheless, the most frequently altered gene in tumor tissue as well as pre-operative and post-operative blood samples was TP53, with a frequency of 43.40%, 23.73%, and 17.07%, respectively. Missense variants were the predominant type of mutations observed in our cohort. The mutation frequencies of genes in our cohort were then further compared with the TCGA stomach adenocarcinoma (STAD) cohorts (Figure S2A). No obvious concordance was observed in genomic alterations between our cohort and TCGA cohort, which might be attributed to the limited number of detectable genes and different cohort backgrounds. The mutational frequency and average allele frequency (AF) both demonstrated a direct proportion total AF per patients, irrespective of pre-operative and post-operative blood samples or tumor tissue (Figure S2B–G).

2.3 One-month post-operative plasma ctDNA level effectively predicted prognosis

Correlation analysis and survival analysis were then conducted to investigate the potential of ctDNA as a promising biomarker in the prediction of prognosis. No significant differences in clinicopathological characteristics were observed between patients with positive versus negative pre-operative or post-operative ctDNA results (Table 2). Multivariable analysis identified 1-month post-operative ctDNA as an independent factor for both OS and progression-free survival (PFS) (Table 3). Among 41 patients with plasma samples collected and underwent ctDNA detection within 1 month after surgery, five (12.2%) patients had high-level ctDNA, while 36 (87.8%) had low-level ctDNA. A significant reduction of OS was observed in patients with high-level ctDNA compared to those with low-level ctDNA (Figure 2A, hazards ratio [HR] = 5.30, p = 0.0022). Patients with high-level ctDNA tended to have a higher risk of recurrence (Figure 2B, HR = 3.85, p = 0.0110). The level of ctDNA AF in patients who died during follow-up was numerically higher than in those who survived (Figure S3A). Similarly, patients who experienced disease progression during follow-up were more likely to have a higher level of ctDNA AF than those who did not (Figure S3D). There was a slight decrease in ctDNA level in 1-month post-operative plasma samples compared to pre-operative samples from patients with good prognosis (Figure S3B,E). An elevation of ctDNA level was observed in some patients who died or experienced relapse during follow-up (Figure S3C,F). These findings indicated the dynamic change of ctDNA level was influenced to a certain extent by therapeutic interventions such as surgery. Moreover, the periodical monitoring of ctDNA in plasma samples could effectively reflect treatment response and predict prognosis effectively.

2.4 ctDNA combining with tumor markers achieved a good efficacy in prognostic prediction

To evaluate the efficacy of prognostic prediction of ctDNA level, the receiver operating characteristics (ROC) analysis was conducted to predict the outcome of prognosis such as survival or recurrence. When conducting ROC for prognostic prediction of post-operative ctDNA, the “survival” was not a time-dependent factor but a binary factor revealing the survival status: living or dead. In terms of the predictive efficacy for survival reported by ROC, ctDNA from the plasma sample yielded an area under the curve (AUC) of 0.645. At a cut-off value of 1.115%, ctDNA achieved a sensitivity of 50.0% and a specificity of 90.0% (Figure 3A,B). When combining ctDNA and conventional serum tumor markers such as carcinoembryonic antigen (CEA), CA (carbohydrate antigen) 19-9, and CA 72-4, a significantly higher efficacy than ctDNA or any tumor markers alone was observed. Particularly, the combination of ctDNA and tumor markers that patients with GC were often recommended to detect in regular examination during follow-up (CEA, CA19-9, and CA72-4) presented a remarkably higher efficacy with an AUC of 0.940 (p = 0.002). Regarding the predictive efficacy for 3-year-survival status reported by ROC, plasma ctDNA level yielded an AUC of 0.645. A slight increase in AUC after adding pTNM stage or tumor size to the model combining ap01 ctDNA level and tumor markers can be observed in Table S1, indicating that the predictive effect for survival could be further enhanced by incorporating conventional clinicopathological information. Regarding recurrence prediction, net ctDNA demonstrated an unsatisfactory performance with an AUC of 0.591, comparable to three biomarkers alone (Figure S4A). However, the combination of ctDNA and tumor markers achieved an AUC = 0.852 (p = 0.010) (Figure S4C). Fitting analysis revealed that ctDNA levels were nearly independent of serum tumor markers like CEA and CA19-9, whereas it positively correlated to CA72-4 in 1-month post-operative plasma samples (Figure S4G–I). This suggested ctDNA could serve as an independent predictive factor exhibiting greater potential in prognostic prediction, especially when combining with clinicopathological and laboratory information.

2.5 Post-operative ctDNA level moderately reflected stromal PD-L1 and PD-1 expression status

The molecular features of tumor and their microenvironment were predominant factors in determining therapeutic response and disease progression, which often function as indispensable references for clinical decision making. Patients were divided into four groups according to quartile of ctDNA level (Figure 4A). Patients in Q4 had a higher risk of progression and were more likely to have a shorter OS compared to those in Q1 (Figure 4B,C). It was illustrated by subgroup analysis that patients who had positive stromal programmed death-ligand 1 (PD-L1) expression in pathological examination tended to have lower ctDNA level post-operatively (p = 0.046) (Figure 4D). Nevertheless, patients with positive stromal programmed death-1 (PD-1) expression appeared to have higher post-operative ctDNA level (Figure 4E). Among patients with positive stromal PD-L1 or PD-1 expression, the grade showed a numerical but insignificant direct proportion with ctDNA level (Figure 4F,G). These results indicates that ctDNA could potentially provide some medication guidance for the long-term application of immunotherapy via dynamic monitoring. However, due to the limited sample size, these findings had limited statistical meaning and needed further validation in future investigations. In addition, TCGA molecular classification was included in analysis (Figure 4H). Most patients had CIN-type GC. Patients with GS-type were more likely to have larger tumor sizes, represented by the longest diameter, while patients with EBV-positive type exhibited the opposite trend. One patient with extremely high ctDNA level in pre-operative and post-operative plasma samples had MSI-type GC.

2.6 DCAF4L2 mutation status in ctDNA and correlated tumor tissue with poor prognosis

In addition, the specific gene mutation in ctDNA that had strong connection with poor prognosis were identified. It may offer us a novel approach to monitor disease progression by tracking the abundance of certain mutations in the plasma sample. The mutational status of ctDNA and tumor tissue in all enrolled patients as well as in patients with poor prognosis were screened, respectively (Figure 5A). The mutation of 13 genes were detected in both tumor tissue and ctDNA from pre- and post-operative plasma samples in patients with poor prognosis. Among 13 genes, the mutation of four genes (TP53, KRAS, FBXL7 and DCAF4L2) were shared by ctDNA and tumor tissue of all patients. The AF value and mutation frequency of these four genes from different sample along with mutation status in TCGA STAD cohort were displayed (Figure 5B and Figure S5A). The external cohort from TCGA database was utilized to explore the potential gene mutation related to poor prognosis. DCAF4L2 was selected as the candidate gene due to its mutation being associated with shorter OS in the pan-cancer cohort, esophageal/GC cohort, and only GC cohort (Figure 5C). In our cohort, post-operative plasma samples from patients with DCAF4L2 alterations showed slightly higher ctDNA levels compared to those without alterations (Figure S5B). Subgroup analysis was then applied in TCGA STAD cohort to discover the mechanism behind. Patients with DCAF4L2 alteration in tumor tissue occupied a larger proportion of MSI-type GC with a significantly higher average MSI MANTIS Score compared to those unaltered ones, which was in consistent with the pattern in our cohort (Figure 5D,E and Figure S5C). Unfortunately, restricted to a small sample size, our results failed to exhibit the significance due to only five DCAF4L2-altered patients in our cohort. MLH1 silencing, a well-acknowledged essential event in MSI GC that was mainly caused by methylated modification, was found to occur more frequently in patients with DCAF4L2 mutations (Figure 5F). To confirm the methylation status of MLH1 in MSI-type GC in the TCGA STAD cohort, CpG island methylator phenotype data were derived and analyzed (Figure S5D). The methylation level of MHL1 was higher in DCAF4L2-altered patients than unaltered ones in several detections via different probes (Figure S5E–I). The Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis showed that multiple common pathways in GC, such as neuroactive ligand–receptor interaction, WNT signaling pathway, proteoglycans in cancer, and cell adhesion molecules, were top enriched, which provides biological explanations on the mechanism of DCAF4L2 mutation exacerbating the disease (Figure S5J). Although the DCAF4L2 mutation can identify patients with poor prognosis, its contribution to the predictive efficacy is minimal (Table S1). This implied that the prognostic prediction could be achieved more accurately and promptly through detecting specific gene mutations in ctDNA from plasma samples.

4.1 Study design and patient cohort

This prospective study enrolled patients with GC from June 2019 to January 2020 in the GC Center of Zhongshan Hospital, Fudan University (Figure S1A). Patients were diagnosed according to standardized diagnostic criteria and clinical guidelines.⁵ Radical resection for GC was performed for the included patients. Patients were scheduled for periodical staging computed tomography of chest, abdomen, and pelvis during the follow-up to exclude metastatic disease and relapse. Plasma samples and paired tumor tissue were also collected according to the prescribed procedure presented in the diagram (Figure S1B). This study is strictly in accordance with the Declaration of Helsinki and was approved by the Research Ethics Committees of Zhongshan Hospital (No. B2018-264). Written informed consent was obtained from each patient.

4.2 Eligibility criteria

The inclusion criteria were as follows: (1) patients with pathologically confirmed adenocarcinoma in stomach or gastroesophageal junction and (2) age above 20 years.

The exclusion criteria were as follows: (1) cancer of other origins than stomach, such as melanoma, (2) patients who had received pre-operative chemotherapy, (3) patients with radiological diagnosis of distal metastasis before surgery, or (4) patients with unresectable GC.

4.3 Sample collection and ctDNA extraction

Blood samples for ctDNA analysis and tumor marker detection were collected pre-operatively (baseline) and post-operatively during follow-up, including the first month after surgery and every 3 months after that. At each time point, 10 mL of blood was drawn into a Roche Cell-Free DNA Collection tube (Roche Diagnostics), and plasma was separated from the cells by centrifugation (1600 × g for 10 min at 4°C) followed by a second centrifugation of the supernatant at 16,000 × g for 10 min at 4°C to remove all cell debris. Plasma samples were frozen at −80°C until further processing in one batch.

4.4 Tumor tissue mutational analysis and ctDNA analysis

The AVENIO Tumor Tissue and ctDNA Surveillance Kit (Roche) covering 197 cancer-related genes were used for analyses. Formalin-fixed, paraffin-embedded (FFPE) tumor tissue from the surgical specimen was analyzed for somatic mutations by next-generation sequencing (NGS). Tumor sequencing was performed with 10–50 ng of FFPE DNA. The cfDNA from plasma was isolated using the Roche AVENIO ctDNA Surveillance Kit (Roche) according to the manufacturer's protocol. Sequencing was performed using 10–50 ng of cfDNA. Median de-duped sequencing depths for cfDNA and FFPE DNA were 5360X and 2099X, respectively. Variants were identified by the AVENIO ctDNA Analysis Software that incorporates bioinformatics methods from CAPP-Seq2 and iDES3 (integrated digital error suppression) to remove PCR duplicates and stereotypical errors from technical artifacts. Somatic variants were derived after filtering out common germline variants in public databases such as ExAC, dbSNP, and 1000 Genomes. AF for a given somatic mutation was calculated as the number of deduplicated reads with that mutation divided by the total number of deduplicated reads covering that genomic position. The validity of the lower limit of detectable AF of 0.2%. Only nonsynonymous single nucleotide variants (SNVs) or small insertions/deletions with an AF of at least 0.2% and 5% were included in further analyses for cfDNA and FFPE, respectively.

For further ctDNA quality control analysis, AF percentage was defined as the representative index of ctDNA level. The ctDNA results with an AF ≥ 40% were excluded for the abatement of clonal hematopoiesis with indeterminate potential. The mutation below the detection threshold presented in the detection report was considered as ctDNA negative, while the positive result was determined as ctDNA positive. Patients were divided into subgroups of high or low level of ctDNA. The best cut-off values of ctDNA at different phase for OS and PFS were calculated by X-tile software and R packages “survival” and “survminer”. In detail, the best cut-off values of pre-operative ctDNA were 1.38% and 1.34% for OS and PFS, respectively. The best cut-off values of post-operative ctDNA for OS and PFS were both 2.54%.

4.5 Histopathological assessment

Pathological characteristics were collected, including tumor location, size, differentiation, Lauren type, and LVI. The molecular classification of GC in this study is based on TCGA standard.⁴¹ Specifically, EBV status was examined by EBV-encoded RNA in situ hybridization. Representative proteins in mismatch repair system, such as MLH1, MSH2, MSH6, and PMS2, were detected by immunohistochemistry (IHC) for determination of MSI status. Any of the proteins detected negative would be determined as mismatch repair deficiency (dMMR), while all proteins positive would be determined as mismatch repair proficiency. The dMMR was classified as the MSI-type GC according to TCGA classification. GS GCs were defined as MSI or EBV-negative diffuse-type GCs. CIN GCs were identified if any of the classification standards above could not be met.

The expression level of PD-1 and PD-L1 in tumor and stroma tissue was determined via IHC assays (antibody: PD-L1 clone 22C3, M3653, DAKO; PD-1, ab137132, Abcam) by experienced pathologists.

4.6 Analysis based on the TCGA database

The primary data of genomic alterations in the TCGA STAD cohort (431 samples) was derived from the TCGA database (https://portal.gdc.cancer.gov) with the information of single nucleotide polymorphism. The visualization of mutations in the TCGA STAD cohort was conducted using the “maftools” package of R. The information on overall survival in different cohorts and relative expression levels of specific genes in both adjacent tissue and tumor were collected from the TCGA database. The data OncoPrint, clinicopathological characteristics, and methylation ChIP-Seq (Chromatin Immunoprecipitation sequencing) of TCGA STAD were obtained from cBioPortal for Cancer Genomics (https://www.cbioportal.org) to validate the relevance between DCAF4L2 mutation and MSI status. The survival data of the external cohort were also derived from cBioPortal: Pan-cancer analysis of whole genomes (ICGC/TCGA, Nature 2020), esophageal and gastric carcinoma (Nature 2017; PanCancer Atlas; Firehose Legacy; Nature 2014), and gastric carcinoma (PanCancer Atlas; Firehose Legacy; Nature 2014). The bioinformatics analysis of the KEGG pathway was conducted using the packages of R. The gene alteration status in TCGA dataset was defined as a general concept including amplification, deletion, and gain or loss of function.

4.7 Statistical analysis

The prognosis outcomes of the study were the OS and PFS of patients with different pre-operative and post-operative plasma ctDNA levels. The distribution of clinicopathological characteristics was compared between two groups by the Wilcoxon rank-sum test. Pearson's chi-square test was used to test the associations between the presence of ctDNA and clinicopathological information. Student's t-test was used to assess the difference of ctDNA level between subgroups of plasma samples. Survival analyses were conducted using the Kaplan–Meier method. The cut-off values were calculated, and survival plots with p-value were calculated using the packages of R. Univariate Cox regression analysis was carried out to explore which parameters were associated with OS and PFS. Variables and parameters with p-value < 0.02 were generally considered to indicate a poor prognosis, such as pTNM and LVI. These could further enter multivariate analysis. Spearman correlation coefficients were determined for the association between ctDNA AF and tumor markers. Visualizations of TCGA gastric adenocarcinoma cohort mutations were carried out in R. In this study, the following software was utilized for statistical analyses: IBM SPSS version 27.0, R Statistics version 4.3.1 (R Foundation for Statistical Computing), and GraphPad Prism version 8.0.0. All hypothesis tests were two-sided. p-value < 0.05 was considered statistically significant.