The MEDLINE database was searched online in March, 2013 for bibliographic information regarding studies on TS expression status and survival in patients with pancreatic adenocarcinoma. The following MESH headings, keywords and text words were used: i) pancreatic and neoplasm OR cancer OR carcinoma OR adenocarcinoma; and ii) thymidylate synthase OR TS. The reference lists of the retrieved articles and previous narrative reviews were scanned for additional potentially relevant articles.

Studies were included in the meta-analysis if they met the following criteria: i) expression of TS was evaluated in the tumor tissue as opposed to serum or metastatic tissue, or in tissue adjacent to the tumour; ii) the expression of TS protein was measured by immunohistochemistry (IHC) or enzyme-linked immunosorbent assay (ELISA); iii) analysis of the association between TS expression and OS; and iv) studies were published as full-text in English. Manual selection of relevant studies was performed based on the summary analysis.

Reviews and conference abstracts were excluded due to limited data for evaluation. When multiple articles pertained to overlapping populations of patients, only the newest, largest, or most informative single article was selected.

Data was carefully extracted independently by two investigators (Guo and Zhu) and controversies were resolved through discussion according to the abovementioned criteria. If an agreement could not be reached, an expert was invited to the discussion. The following information was extracted from each study: first author, year of publication, method, specimen, number of eligible patients, antibody, resection and OS results.

The effect of TS expression on OS was estimated for each study by the hazard ratio (HR) with its 95% confidence interval (CI), according to the methods described by Tierney et al (14). In summary, when the estimated HR and its standard error were described in the publications, we obtained these values directly; when these statistical variables were not provided explicitly in an article, they were calculated directly using two of the following parameters: the CI for the HR, the log-rank statistic, its P-value or the O-E statistic (difference between numbers of observed and expected events); when those data were not available, the following were investigated: total number of events, number of patients at risk in each group and the log-rank statistic or its P-value, allowing calculation of an approximation of the HR estimate; when the only available data were in the form of graphical representations, they were calculated from Kaplan-Meier survival curves. The Kaplan-Meier curves were read by two investigators using Engauge Digitizer software, version 4.1 (Mark Mitchell, Boston, MA, USA) independently to reduce inaccuracy in extracted survival rates.

A χ² test-based Q statistic test for between-study heterogeneity was used with a P-value of 0.1 rather than 0.05 to determine statistical significance (15). P>0.10 indicated a lack of heterogeneity among studies, so the fixed-effects model was used to combine HR. Otherwise, the random-effects model was used (16). By convention, the effect of TS on OS was considered as statistically significant if the 95% CI for the overall HR did not overlap 1. Evidence of publication bias was evaluated by the funnel plot with Begg's test (17) and Egger's linear regression asymmetry test (18). For these analyses, P<0.05 was considered representative of statistically significant publication bias. All the statistical analyses were performed by STATA 12.0 software (StataCorp, College Station, TX, USA).

The characteristics of the eligible studies are summarized in Table I. A total of 5 studies (9–11,13,19), published between 2002 and 2012, met the inclusion criteria for this review. A total of 425 patients were subjected to the final analysis. All the studies were based on the data of retrospective analysis. Among the 5 studies investigating OS, a significant association between TS high expression and OS was identified in 3, including 2 associating TS high expression with better OS and 1 associating TS high expression with worse OS. The remaining 2 studies yielded negative results.

The expression of TS was measured by IHC in 4 of the 5 studies and by ELISA in the remaining study. The IHC techniques used varied widely among the studies, with a wide range of dilutions (1:15–1:1,000). According to the cut-off values for TS high expression as defined by the authors of each study, a total of 230 patients (54.1%) in this meta-analysis exhibited high TS expression, with a range of 25–75%.

Three studies accrued patients who had received R0 and R1 resection (10,13,19) and stratified results regarding the association between TS expression with OS for resection type (R0 or R1) were only mentioned in 1 study (19). Only the patients who underwent R0 resection were analyzed in the remaining 2 studies (9,11).

The pooled HR was 0.72 (95% CI: 0.41–1.25; Z=1.18, P=0.238), indicating that TS expression exerted no significant survival effect on patients with pancreatic adenocarcinoma. Heterogeneity testing revealed that there was inter-study heterogeneity (Q=20.46, P<0.001). (Fig. 1). When we limited the analysis to the studies assessing R0 resection patients, the combined HR was 0.46 (95% CI: 0.31–0.68; Z=3.95, P<0.001), without any evidence of heterogeneity (Q=0.08, P=0.961), which indicated that a high expression of TS was significantly correlated with better OS in patients with pancreatic adenocarcinoma who underwent R0 resection (Fig. 2).

Publication bias statistics obtained by the methods of Begg and Mazumdar (17) and Egger et al (18) were as follows: all 5 eligible studies (P=0.806 and P=0.414); and the 3 studies assessing R0 resection (P=1.00 and P=0.279). This suggested absence of publication bias in all the studies.