Association Between Liquid Biopsy and Prognosis of Gastric Cancer Patients: A Systematic Review and Meta-Analysis

Background: Reports regarding liquid biopsy and gastric cancer (GC) have emerged rapidly in recent decades, yet their prognostic value still remains controversial. This study was aimed to assess the impact of liquid biopsy, including circulating tumor cells (CTCs) and cell-free nucleic acids, on GC patients' prognosis. Methods: PubMed, Medline, EMBASE, and ClinicalTrial.gov databases were searched for studies that report GC patient survival data stratified by CTC/circulating tumor DNA (ctDNA)/circulating miRNAs' status. The hazard ratios (HRs) and their 95% confidence intervals (CIs) for patients' overall survival (OS) and disease-free survival (DFS)/progression-free survival (PFS) were recorded or calculated depending on circulating target status. Results: We initially identified 4,221 studies, from which 43 were eligible for further analysis, comprising 3,814 GC patients. Pooled analyses showed that detection of certain CTCs, ctDNA, and circulating miRNA was associated with poorer OS (CTCs: HR = 1.84, 95%CI 1.50–2.26, p < 0.001; ctDNA: HR = 1.78, 95%CI 1.36–2.34, p < 0.001; circulating miRNA: HR = 1.74, 95%CI 1.13–2.69, p < 0.001) and DFS/PFS (CTCs: HR = 3.39, 95%CI 2.21–5.20, p < 0.001; ctDNA: HR = 2.38, 95%CI 1.31–4.32, p = 0.004; circulating miRNA: HR = 3.30, 95%CI 2.39–4.55, p < 0.001) of GC patients, regardless of disease stage and time point at which sample is taken (at baseline or post-treatment). Conclusions: The presence of CTCs and/or cellular components identifies a group of GC with poorer prognosis. Among circulating markers, CTCs demonstrated a stronger and more stable predictive value for late-stage disease and among Mongolian populations with GC. Less data are available for ctDNA and miRNA; however, their presence may also reflect aggressive biology and warrants further prospective study.


INTRODUCTION
Gastric cancer (GC) remains the fifth most common cancer and the third leading cause of cancer-related death worldwide (1,2). Although some therapeutic advances have been made, its prognosis remains unfavorable owing to the aggressive tumor biology, late detection, and high disease progression/recurrence rate (3). Few clinicopathological factors are used to guide therapy or disease monitoring, and ideal peripheral blood biomarkers have been lacking. Although enhanced endoscopic techniques, such chromoendoscopy (4) and endoscopy with narrow-band imaging (NBI) (5), are considered to be the more reliable and credible methods for diagnosis of GC than conventional diagnostic tools, their applications are limited because of their invasive nature and cost-efficacy concerns (5).
Theoretically, tumor-derived blood-based biomarker tests have multiple application in GC including detecting/monitoring response after therapies, identification of actionable tumor alterations, and patient stratification (17,18). Currently, the diagnostic value of liquid biopsy is still under debate, and it has been questioned for its low sensitivity and yields in some series (12,19). In contrast, the prognostic importance is increasingly supported by mounting evidence in breast (20) and colorectal (21) cancers. Although cfNAs include several cellular components, the most commonly investigated in GC research are circulating tumor deoxyribonucleic acid (ctDNA) (22) and circulating microRNA (miRNA) (23). Variability in detection methodology, genomic coverage, specimen processing, and reproducibility has not always been consistent. Moreover, the most appropriate sampling time point for accurate detection (at baseline or post-treatment), the most appropriate test population and disease stage, and even the predictive value of certain types of biomarkers have not yet been agreed (12,24). With the continuously emerging data in GC, there is a need to conduct quantitative analysis evaluating the most commonly used liquid biopsy methods currently in GC management. Therefore, we sought to conduct a systematic review and meta-analysis to evaluate the significance of CTCs and cfNAs in predicting GC progression and recurrence in a methodologically consistent manner.

Literature Search
MOOSE (Meta-analysis Of Observational Studies in Epidemiology) (25) and PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) (26) guidelines were applied to conduct the systematic review. The following databases were systematically searched for relevant studies published up to December 2017: PubMed, Medline, EMBASE, Web of Science, and the Cochrane Central Register of Controlled Trials. Bibliographies of all relevant papers were also checked for further eligible studies. There was no restriction on language of publication (Table S1).

Selection Criteria
Studies were included in the analysis if they met the following criteria: (1) they enrolled patients with pathologically confirmed gastric or gastroesophageal junction adenocarcinoma; (2) they reported GC patient survival data stratified by CTC/ctDNA/circulating miRNA status (presence/positive and absence/negative); (3) they provided sufficient data for determining or calculating a hazard ratio (HR) and 95% confidence interval (95%CI); and (4) they enrolled patients who did not overlap with patients included in other eligible studies.
Studies were excluded if (1) fewer than 20 patients were analyzed; (2) samples were not drawn from peripheral blood (e.g., from urine or bone marrow); or (3) the histology type of included GC patients was squamous carcinoma or neuroendocrine carcinoma.

Data Extraction
Two authors (HX and JC) independently reviewed the eligible studies and extracted the following information: first author name, publication year, number of patients analyzed, age, gender, tumor stage, clinical treatment, volume and timing of blood withdrawal, marker detection method, cutoff value, positive ratio, and follow-up duration, if provided. When more than one marker was assessed in studies and an HR for survival or the survival curve was provided for each marker, results for all these markers were recorded as independent data sets.

Assessment of Risk of Bias
Risk of bias for individual studies was assessed using a modified Cochrane risk of bias instrument that included evaluation options of "definitely or probably yes" or "definitely or probably no" or "unknown or unclear" (27). The items included "adequate eligibility, " "the measurement equality, " "controlled confounding, " "adequate follow-up, " "free of selective outcomes, " and "other factors" (Table S2) (21).

Statistical Analysis
The HRs and their 95%CIs for overall survival (OS) and disease-free survival (DFS)/progression-free survival (PFS) were recorded. For studies where HRs were not provided, we approximated HRs from the Kaplan-Meier curves with the use of an HR calculation Excel spreadsheet provided by Tierney et al. (28). All HR data extraction and calculations were performed independently by YHG and HQX, and disagreements were resolved by discussion. Survival outcomes generated using multivariate analysis models were preferentially used if available, to ensure results are as clinically relevant as possible. By convention, an HR > 1 implies a worse prognosis in the circulating marker positive/upregulated group than in the negative/downregulated group, and p < 0.05 indicated statistical significance.
We pooled the extracted HRs using the generic inverse variance method. We anticipated interstudy heterogeneity and so used a random-effect analysis model preferentially (29). If no obvious heterogeneity was observed (p > 0.05), then a fixed-effect model was applied. Analyses were conducted using Stata 12.0 (StatCorp, College Station, TX, USA).

Sensitivity Analysis, Subgroup Analysis, and Meta-Regression Analysis
The stability of pooled HRs was tested by one-way sensitivity analysis with omission of a single study. Subgroup analyses and meta-regression were performed to explore potential sources of heterogeneity, and the following clinicopathological features were stratified: sampling time (at baseline or postoperatively), number of tested targets, cutoff value, tumor-node-metastasis (TNM) stage, risk of bias level, statistical methodology employed, ethnicity, and sample size. Any subgroup comprising fewer than two studies was excluded from the analysis.

Baseline Study Characteristics
Forty-three studies were eligible for inclusion, comprising 3,814 patients. These included 20 studies reporting on CTCs, 10 on ctDNA, and 13 on circulating miRNAs. Considering CTCs could also be performed at the DNA or RNA (mRNA or microRNA) level, we classified enrolled studies into relevant groups according to the authors' description in their report (Figure 1).
The baseline characteristics and study design variables of the included studies are shown in Table 1. All studies were written in English. Sample sizes ranged from 27 to 277 patients (median: 73 patients). The studies were conducted in 11 countries or regions (China, Egypt, Germany, Greece, Hong Kong, Italy, Japan, Poland, Spain, Taiwan, and Thailand).
All 43 studies applied a molecular or cytological detection method analyzing venous blood [polymerase chain reaction (PCR), quantitative reverse PCR (qRT-PCR), methylationspecific PCR (MSP), quantitative MSP (qMSP), next-generation sequencing (NGS), immunofluorescence (IF), CellSearch System, or colorimetric membrane array (CMA)]. Notably, three studies applied a combination of molecular and cytological detection methods (35,37,39). Five studies (37,43,57,71,72) analyzed the same patient cohort but using two different targets. To account for this, both markers were included in the pooled analysis, whereas the total number of patients was only counted once. The assessment of risk of bias for individual studies showed 31 and 12 studies with a low risk of bias and a high risk of bias, respectively. HRs for OS and DFS/relapse-free survival (RFS) could be extracted from 35 to 16 studies, respectively. Publication bias analyses were carried out for the analysis of all studies in Egger's and Begg's tests on OS and DFS/RFS, but no relevant publication bias was observed ( Figure S1).

Circulating Tumor Cells
HRs for OS were available in 17 studies, representing 1,239 patients. Two HR estimates for OS were extracted from Uen et al. for the reason mentioned in the Methods part (43). The pooled HR showed a significant prognostic effect of CTC detection in GC patients (HR = 1.84, 95%CI 1.50-2.26, p < 0.001, Figure 2A), with moderate heterogeneity (I 2 = 44%, p = 0.024).
HRs for DFS/PFS were available in 11 studies, representing 848 patients. The pooled HR showed a significantly increased risk of disease progression or recurrence in patients with CTC positivity (HR = 3.39, 95%CI 2.21-5.20, p < 0.001). The heterogeneity between studies was significant (I 2 = 63.8%, p = 0.002).
Sensitivity analyses conducted by omitting each single study changed this result only marginally ( Figure S2). Table 2  Meta-regression identified cancer stage and patient ethnicity as variables influencing OS HR estimates for CTCs ( Table 2, p = 0.010 and p = 0.008, respectively). The presence of CTCs is associated with a higher HR for OS in studies enrolling only latestage patients (HR = 2.81, 95%CI 1.79-4.40, p < 0.001) than studies enrolling with both early-and late-stage patients (HR = 1.84, 95%CI 1.50-2.26, p < 0.001). Nevertheless, both results from subgroups by TNM stage indicated a significant association between CTCs presence and worse prognosis of GC patients.
Studies involving GC patients of Mongolian ethnicity had a significantly higher pooled HR (2.04, 95%CI 1.64-2.54, p < 0.001) than had studies involving Caucasian patients (HR = 1.34, 95%CI 1.16-1.54, p < 0.001). This was further supported by tests for interaction (p = 0.008, Table 2). However, these differences by disease stage and ethnicity found in a subgroup analysis of OS HRs were absent in the analysis of DFS/PFS ( Table 2). No other variables were found to be significant, which may be because of the relatively limited number of studies reporting DFS/PFS (11 studies in total, only one of which studied a primarily Caucasian patient population).
The subgroup analysis on sampling time showed a prognostic effect of CTC detection for both time points (baseline and during/post-treatment). HRs for CTCs predicting the survival of GC patients where liquid biopsies were taken during/posttreatment were higher than HRs of patients where biopsies were taken at baseline. This was the case for both OS (HR = 2.30, 95%CI 1.52-3.49, p < 0.001 during/post-treatment; HR = 1.63, 95%CI 1.30-2.04, p < 0.001 at baseline) and DFS/PFS (HR = 4.04, 95%CI 1.21-13.44, p = 0.023 during/post-treatment; HR = 3.15, 95%CI 1.99-5.0, p < 0.001 at baseline). However, this difference did not reach statistical significance and could not be substantiated by further tests of interaction.

Circulating Tumor DNA
HRs for OS were reported in six studies, representing 624 patients. More than one HR for OS was extracted from three studies, because multiple detection approaches were used. The pooled HRs showed a significant prognostic effect of the detection of ctDNA in GC patients' OS (HR = 1.78, 95%CI 1.36-2.34, p < 0.001, Figure 3A), with moderate heterogeneity (I 2 = 46.7%, p = 0.059). No ctDNA targets were assessed by more than two independent studies. Therefore, a subgroup analysis by target was not performed. A subgroup analysis by other variables revealed that ctDNA presence was significantly associated with shorter survival for all subgroups except in studies conducted primarily in Caucasian patients (N = 2, HR = 1.55, 95%CI 0.85-2.83, p = 0.156). However, this result must be interpreted with caution, given the small sample size. A Galbraith plot indicated that the study by Pimson    decrease in heterogeneity (I 2 = 33.2%, p = 0.163), but the association between ctDNA and OS remained significant (HR = 1.64, 95%CI 1.28-2.10, p < 0.001). HRs for DFS/PFS were reported by five studies utilizing ctDNA, representing 731 patients. The pooled HR showed a significantly increased risk of disease progression or recurrence in patients with ctDNA detection (HR = 2.38, 95%CI 1.31-4.32, p = 0.004). Heterogeneity between studies was significant (I 2 = 87.2%, p < 0.001). A Galbraith plot and a sensitivity analysis were performed to explore the source of heterogeneity and stability of the results. Although a sensitivity analysis showed that omission of any single study would not substantially alter the outcomes, the Galbraith plot showed that Fang et al. (59) and Ling et al. (53) were outliers and the main contributors to heterogeneity ( Figure S3B). Excluding these two studies reduced heterogeneity somewhat (I 2 = 56.5%, p = 0.100) and made the association between ctDNA and DFS/PFS more significant (HR = 2.19, 95%CI 1.31-3.66, p = 0.003, Figure 3B).
After sensitivity analyses were performed, it was found that by excluding the only two studies with an HR estimate < 1 (67,72), the adjusted pooled HR for OS was higher (HR = 2.13, 95%CI 1.61-2.83, p < 0.001) whereas heterogeneity was substantially reduced (I 2 = 53.4%, p = 0.014). In a manual review of the original work of these two studies, the authors considered the two targets, miR-203 and miR-122, as anti-tumor   microRNAs on the basis of biological function. Therefore, a subgroup analysis and a meta-regression analysis were performed after excluding these targets, and then significant associations between circulating miRNA detection and OS were found in both sample time point groups (baseline and during/post-treatment). Unlike CTC analyses, a subgroup analysis stratified by tumor stage (all stages vs. advanced stage only) and ethnicity (Caucasian vs. Mongolian) did not alter the bias and differences significantly between these subgroups. HRs for DFS/PFS in circulating miRNA were available in five studies, representing 584 patients. The pooled HR for DFS/PFS was 3.30 (95%CI 2.39-4.55, p < 0.001, Figure 4B). Heterogeneity between HR estimates was not significant (I 2 = 0.0%, p = 0.835).

DISCUSSIONS
Here, we report the largest meta-analysis of circulating tumorderived biomarkers and identify prognostic value for CTCs. Our meta-analysis provides strong evidence, even after adjustment for clinical variables. With over 3,800 included GC patients, our study is the most comprehensive systematic review of the association between liquid biopsy and GC prognosis to date, substantially larger than previous studies (73,74).
Importantly, we attempted to address biomarker detection method, study heterogeneity, and disease stage. The association between biomarker detection (CTC, ctDNA, or miRNA) was relatively stable and not influenced largely by liquid biopsy detection methods or disease stage. Even among GC patients where samples were taken post-treatment, the association remained significant, highlighting the potential clinical utility of blood-based biomarkers in GC.
Overall, we observed a stronger association between circulating marker detection and DFS than OS, suggesting an important role in prognosis and patient stratification, particularly in non-metastatic patients. We acknowledge that an optimal cutoff value for each detection method remains to be determined, and a decreased heterogeneity was observed in the receiver operating characteristic (ROC) cutoff determination subgroups, indicating more consistent results in studies that adapted ROC curves to determine patients' tumor status.
Among the detection platforms examined, several important observations warrant further discussion, and the analysis of whole CTCs can be performed at the DNA or RNA (mRNA or microRNA) and protein levels, whereas the analysis of ctDNA and microRNAs can be performed only at the genomic level. For example, one alternative to enumerating CTCs by immunocytochemistry (ICC) is to estimate their presence using RT-PCR to discover epithelial transcripts, which should not be present in normal hematopoietic cells. However, detection of CTCs often requires more cumbersome enrichment and detection methods, whereas the detection of cfNAs can be performed using blood plasma or serum, and easier methods (24). Huang et al. (73) and our previous report (75) have demonstrated the significance of CTC and ctDNA in GC patients' prognosis prediction.
Among studies examining CTCs, the CellSearch System was the most widely used method for detecting and enumerating CTCs from blood samples, using a combination of epithelial markers (EpCAM+; cytokeratin 8, 18, and/or 19; and CD45-). It is still the first and only actionable commercial test for detecting CTCs in cancer patients, including metastatic breast (14), prostate (15), and colorectal cancers (16). Our results further support its application for GC patients as a statistically significant predictor of shorter OS and DFS/PFS.
Another popular marker type in CTC detection was found to be cytokeratins (CKs). CKs have been found to have different predictive values in patients from Asian (N = 2, HR = 3.54, 95%CI 1.84-6.82, p < 0.001) and Western populations (N = 3, HR = 1.38, 95%CI 0.73-2.61, p = 0.328), which suggests that they may play a different role in different ethnicities. Moreover, CKs have tended to serve as biomarkers in a combination of their own components (e.g., CK18, CK19, and CK20) (76) or alongside other targets such as EpCAM (77,78) to identify the epithelial cells more precisely.
Circulating tumor DNA is composed of small fragments of nucleic acid that are not associated with cells or cell fragments (79). The most widely used method of detection is methylated DNA in plasma/serum, which is usually identified by MSP or quantitative MSP (qPCR) assays (80). All included studies withdrew blood for ctDNA detection at the baseline time point, which is probably because ctDNA is rapidly cleared from circulation after surgery or other therapy because of its short half-life (80). However, a previous study reported that DNA methylation is relatively chemically stable and can be detected at a sensitivity of up to 1:1,000 molecules (81). It is therefore not surprising that most of the studies included in our review focused on epigenetic regulation of circulating markers. Only Fang et al. investigated the role of gene mutation and copy number.
Although the dysregulation of ctDNA is relatively common in gastroesophageal cancers (22), a reliably detectable prognostic ctDNAs with high specificity is yet to be identified. Our metaanalysis only covers genes and epigenetic regulators relevant to GC. Whole gene screening assays, especially for genetic mutations, are required to identify more associations (82,83).
Beyond CTCs and ctDNA, circulating miRNAs (miRNAs) are a large group of short, non-coding RNAs, 19-25 nucleotides long, which regulate gene expression by pairing to the 3 ′ untranslated region (3 ′ -UTR) of their target mRNA (84). It has been suggested that miRNAs could function as either tumor suppressor or oncogenes by regulating gene expression at transcriptional and translational levels in GC (85). Notably, although not all detected CTCs are predictors of adverse outcomes (86), the majority of them are. In contrast, certain miRNAs detected in serum/plasma may be positive predictors of GC patient survival, acting as tumor-suppresser genes, such as miR-192 and miR-203. Nevertheless, our results also support previous evidence that oncogenic circulating miRNAs are strongly significant predictors of poorer outcomes, particularly for GC recurrence, and progression (HR = 3.41, 95%CI 2.48-4.69, p < 0.001; I 2 = 0.0%, p = 0.670).
In the past, it has been difficult to obtain tumor samples from GC patients without surgery, as endoscopic biopsy provides limited genetic or cellular materials in most cases. Although the optimal platform remains open to debate, the ability of ctDNA to simultaneously detect genomic alterations is attractive and might have a prognostic role. Our meta-analysis supports the use of a series of detection targets and methods for predicting GC patients' OS and DFS. Intriguingly, our data suggest that detection of certain circulating markers at any time, pre-treatment, or post-treatment, provides important prognostic information. Among the overall advanced disease population, the presence of CTCs and tumor-related nucleic acids may help identify those patients that could benefit most, or at least, from systematic therapy including chemotherapy, target therapy, or immunotherapy (87,88). In the era of NGS and a combination of multi-analytic biomarkers (89)(90)(91), our meta-analysis provides a solid foundation and methodological reference for further study.
We acknowledge several limitations to our large metaanalysis. First, studies may tend to selectively report their positive results, leading to risk of selection and publication bias. Second, the majority of our studies enrolled patients from all disease stages, making it difficult to stratify the prognostic value of circulating biomarkers by stage. In addition, a subgroup analysis of some variables involved groups with small sample sizes, which might bias our conclusions. Although meta-regression has indicated that tumor stage and ethnicity may contribute to inter-study heterogeneity in prognostic value, large, multicenter prospective studies based on homogeneous patient populations are still required to validate our findings.

CONCLUSIONS
In conclusion, results of this meta-analysis demonstrated a significant role for liquid biopsy, including CTCs, ctDNA, and circulating miRNA, in predicting worse prognosis of patients with GC. By analyzing currently available studies, CTCs demonstrated a stronger and more stable predicative value in late-stage disease and Mongolian populations compared with early-stage disease and Caucasian populations, respectively. Careful selection of circulating markers and standard detection methods are likely to be fundamental to optimizing the accuracy of liquid biopsy in determining GC patients' prognosis. And further multicentered studies applying specific circulating biomarkers are warranted to clarify the clinical validity of liquid biopsy and its utility in GC patients.

AUTHOR CONTRIBUTIONS
YG, BW, and LC conceived the study. YG, HX, and JCu conducted the literature searches and extracted the data. AC, WS, and JL took part in the analysis and interpreted the data. YG, ZQ, and HX drafted the manuscript. RR, TC, JCh, and SK critically revised the manuscript. HL, JL, and ZQ helped in making the figures and tables. BW and KZ double-checked the extracted data.