The meta-analysis was designed and conducted based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) reporting guidelines, which are considered the gold standard for reporting systematic reviews and meta-analyses. The specific search strategy is Tertiary Lymphoid Structures OR Lymphoid Structure, Tertiary OR Lymphoid Structures, Tertiary OR Tertiary Lymphoid Structure OR Ectopic Lymphoid-Like Structures OR Lymphoid-Like Structure, Ectopic OR Ectopic Lymphoid Tissues. The protocol for this meta-analysis study can be found in PROSPERO (19).

The eligible studies should meet the following criteria (1): GI cancers confirmed by pathological diagnosis (2); Detection of the expression levels of TLS in human tumor tissues (3); Providing survival data, including hazard ratio (HR) and 95% confidence interval (CI) measurements for OS, RFS, PFS, or PFS, or providing Kaplan-Meier curves based on TLS grouping (4); Providing the methods for TLS detection and evaluation. The following studies have been excluded from consideration due to various reasons (1): Comment, animal studies, letter, edit, reviews and meta-analysis (2); Conducting multiple studies using the same set of samples or participants (3); No insufficient data or no prognostic information.

Extract the following data from the included studies: year of publication, region, sample size, sex, cancer type, TLS cut-off criteria, follow-up time(months), survival analysis (OS, RFS, DFS or PFS), HR and 95%CI. If HR and 95% CI are not provided, Engauge Digitizer software version 4.1 was used to plot the Kaplan–Meier curves and extract the multiple survival rates to estimate the HRs and 95% CIs (20). Quality assessment was performed using the Newcastle–Ottawa quality assessment scale (NOS). NOS criteria scores range from 0 (lowest) to 9 (highest), and a NOS score ≥6 is considered a high-quality study. Two reviewers (Kaile Zhang and Yule Yang) independently assessed the quality of the eligible studies and extracted the data, and any disagreement was resolved through discussion with the third (Menghan Cao) (21).

In this study, the gene expression and clinical information of gastrointestinal cancer patients were downloaded from the TCGA database (https://portal.gdc.cancer.gov/). The patients were divided into two groups based on the scores of 9 TLS-related genes (CCR6, CD1D, CD79B, CETP, EIF1AY, LAT, PTGDS, RBP5, and SKAP1), namely the TLSscore high group (upper tertile) and the TLSscore low group (lower tertile) (22). The ESTIMATE algorithm was used to analyze the immune, stromal, and ESTIMATE scores. The differences in survival between the two groups were compared using the logarithmic rank test, and visualized using Kaplan-Meier.

The proportion of 28 immune cells in the tumor microenvironment is determined using the single-sample gene set enrichment analysis (ssGSEA) method.

The statistical analysis was performed using Stata 15.0. It involved calculating the correlations between TLS and OS, RFS, PFS, and DFS. If P<0.05 and I² >50%, it indicated high heterogeneity, and a random-effects model was applied. Otherwise, a fixed-effects model was used. Egger’s and Begg’s tests were employed to assess publication bias. If significant publication bias was detected, the trim and fill method was utilized to adjust the results. Additionally, a sensitivity analysis was conducted by systematically excluding individual studies in order to evaluate the robustness of the meta-analysis. P<0.05 was considered statistically significant.

After the initial search, we eliminated a total of 4435 duplicate articles. Then, we carefully read the titles and abstracts of the remaining articles and excluded 3286 of them. Subsequently, we obtained the full text of the remaining 64 articles and conducted a thorough evaluation. Finally, we selected and included 25 articles for our study (12–18, 23–40). These 25 articles encompassed 32 individual studies and involved a total of 5778 patients. The PRISMA flowchart depicting the entire selection process can be seen in the provided ( Figure 1 ).

The included studies in this research are summarized in ( Table 1 ), which consists of 8 studies on gastric cancer (GC), 6 studies on colorectal cancer (CRC), 7 studies on hepatocellular carcinoma (HCC), 4 studies on esophageal cancer (EC), 4 studies on pancreatic cancer (PC), and 3 studies on intrahepatic cholangiocarcinoma (ICC). Among these studies, 20 were conducted in China, 7 were from Japan, and 2 each were from Finland, the United States, Australia, and Germany. Moreover, 10 studies only provided overall survival (OS) data, 1 study reported only disease-free survival (DFS) data, 1 study solely focused on relapse-free survival (RFS) data, and another study presented progression-free survival (PFS) data exclusively. Interestingly, 11 studies provided both OS and RFS data, while 4 studies included both OS and DFS data, as well as OS and PFS data. The incorporated studies employed four cut-off criteria to designate TLS: namely, Presence, Density, Degree of maturation, and Maximum dimension. The NOS scores of the 32 studies ranged from 6 to 8, signifying an exceptional standard of the encompassed research. Every study embraced pertinent insights regarding TLS within malignant growths in this article.

TLS is divided into TLS-high and TLS-low based on different cut-off criteria. Among the 31 studies included in this paper, different cut-off criteria correspond to different HR. In the subsequent investigation of the relationship between TLS and OS, RFS, PFS, and DFS, we have established inclusion criteria. When a study includes two different TLS cut-off criteria, we prioritize the HR corresponding to Density, Degree of maturity, or Maximal diameter. If a study simultaneously uses Density and Degree of maturity or Density and Maximal diameter as the grading methods for TLS, we select the HR corresponding to Density.

Based on a comprehensive analysis of 29 studies evaluating the association between TLS and OS in GI cancer, it was found that TLS-high is significantly correlated with longer OS(HR=0.525, 95%CI:0.447-0.616, P<0.001)( Figure 2A ). However, it should be noted that this analysis showed significant heterogeneity, and a random effects model was used to account for this(I ^{2 =} 65.7%, P<0.001)( Figure 2A ). Moreover, subgroup analysis based on different cancer types revealed that TLS-high is closely associated with extended OS in GC(HR=0.422, 95%CI:0.283-0.627,P<0.001), HCC(HR=0.532, 95%CI:0.391-0.726,P=0.003), EC(HR=0.393, 95%CI:0.271-0.570,P<0.001), PC(HR=0.390, 95%CI:0.290-0.525,P<0.001), and ICC(HR=0.493, 95%CI:0.421-0.577,P<0.001), while no significant relationship was found between TLS and OS in CRC(HR=0.775, 95%CI:0.570-1.053,P=0.103) ( Figure 2B ). Interestingly, significant heterogeneity was observed in GC(I ^{2 =} 80.6%,P<0.001) and HCC(I ^{2 =} 78.0%,P<0.001) ( Figure 2B ).

Moving on to the assessment of TLS in relation to RFS, DFS, and PFS, 12, 6, and 5 studies were included, respectively. The analysis showed a significant association between TLS-high and extended RFS(HR=0.546, 95%CI:0.461-0.647, P<0.001)( Figure 3A ), DFS(HR=0.519, 95%CI:0.417-0.646, P<0.001)( Figure 3B ), and PFS(HR=0.588, 95%CI:0.406-0.852, P=0.005)( Figure 3C ). Notably, the analysis of RFS(I ^{2 =} 16.8%, P=0.279)( Figure 3A )and DFS I ^{2 =} 5.7%, P=0.380)( Figure 3B ) did not exhibit significant heterogeneity, and a fixed effects model was used, whereas significant heterogeneity was observed in the analysis of PFS[I ^{2 =} 61.2%, P=0.036)]( Figure 3C ), requiring the use of a random effects model.

In addition, two studies assessing the relationship between TLS and OS in the context of using ICIs for treatment were included, as well as three studies evaluating the relationship between TLS and PFS. The study conducted by Mori Y et al. and Hayashi Y et al. included patients who received treatment with anti-PD-1 monoclonal antibodies. On the other hand, the study conducted by Shang T et al. did not explicitly specify the type of ICIs used in their research. The results consistently showed a significant association between TLS-positive and extended OS[HR=0.475, 95%CI:0.282-0.799, P=0.005)]( Figure 4A ) and PFS[HR=0.576, 95%CI:0.381-0.871, P=0.009)]( Figure 4B ). No significant heterogeneity was found in any of these studies, and a fixed effects model was applied(I ^{2 =} 0.0%, P=0.352)( Figure 4A ), (I ^{2 =} 33.4%, P=0.223)( Figure 4B ).

Notably, in previous studies examining the relationship between TLS and OS, significant heterogeneity was still observed in the GC and HCC subgroups. This may be attributed to the use of different TLS cut-off criteria in some studies, resulting in different conclusions. Therefore, to ascertain the potential impact of cut-off criteria on the evaluation of TLS prognosis, studies using different criteria such as Presence, Density, Degree of maturity, and Maximal diameter were included. However, due to the limited number of included studies, further investigation on the impact of cut-off criteria on DFS and PFS was not possible. Firstly, we included 15 studies that used presence as a cut-off criterion to study the relationship between TLS and OS. The evaluation showed a significant correlation between TLS-high and prolonged OS in the included studies(HR=0.590, 95%CI:0.474-0.733, P<0.001)( Figure S1A ). However, there was significant heterogeneity observed, so a random-effects model was used(I ^{2 =} 62.7%, P=0.001)( Figure S1A ). Subgroup analysis based on cancer types revealed no significant association between TLS and OS in HCC(HR=0.850, 95%CI:0.721-1.002, P=0.053)and CRC(HR=0.731, 95%CI:0.417-1.282, P=0.272), while a significant correlation was found in PC (HR=0.351, 95%CI:0.248-0.498, P<0.001)and other tumors(HR=0.629, 95%CI:0.508-0.778, P<0.001)( Figure S1B ). Next, we included 15 studies that used density as a cut-off criterion to study the relationship between TLS and OS. The evaluation showed a significant correlation between TLS positivity and prolonged OS in the included studies(HR= 0.516, 95%CI:0.450-0.591, P<0.001)( Figure S1C ). No significant heterogeneity was found, so a fixed-effects model was used(I ^{2 =} 23.3%, P=0.195)( Figure S1C ). Subgroup analysis based on cancer types showed a close correlation between TLS-high and OS prolongation in GC(HR=0.466, 95%CI:0.302-0.719,P<0.001), HCC(HR=0.401, 95%CI:0.307-0.524, P<0.001),EC(HR=0.393, 95%CI:0.271-0.570,P<0.001),PC(HR=0.510, 95%CI:0.291-0.893,P=0.019) and ICC(HR=0.612, 95%CI:0.479-0.781,P<0.001) ( Figure S1D ). However, no significant relationship was observed in CRC(HR=0.794, 95%CI:0.550-1.145,P=0.217) ( Figure S1D ). Furthermore, after including 3 studies that used maturity as a criteria to study the relationship between TLS and OS, we found a similar significant correlation between TLS-high and prolonged OS(HR=0.222, 95%CI:0.068-0.0730, P=0.013)( Figure S1E ). Due to significant heterogeneity observed in the studies, a random-effects model was used(I ^{2 =} 84.7%, P=0.001)( Figure S1E ). Moving on to the relationship between TLS and RFS, we included 10 studies that used presence as a cut-off criterion. We found a significant correlation between TLS and prolonged RFS in these studies(HR=0.591, 95%CI:0.495-0.705, P<0.001)( Figure S2A ). No significant heterogeneity was observed, so a fixed-effects model was used [(I ^{2 =} 33.5%, P<0.122]( Figure S2A ). Subgroup analysis based on cancer types revealed no significant association between TLS and RFS in CRC(HR=0.874, 95%CI:0.527-1.450, P=0.602)( Figure S2B ), while a significant correlation was found in HCC(HR=0.638, 95%CI:0.518-0.787, P<0.001), PC(HR=0.260, 95%CI:0.137-0.496, P<0.001), and other tumors HR=0.365, 95%CI:0.200-0.665, P=0.001)( Figure S2B ). In the 2 studies that evaluated the relationship between TLS and RFS using density as a cut-off criteria, we found a significant correlation between TLS and RFS(HR=0.457, 95%CI:0.327-0.640, P<0.001)( Figure S2C ). No significant heterogeneity was found in these studies, so a fixed-effects model was used(I ^{2 =} 0.0%, P=0.442)( Figure S2C ).

We found that different studies have used various criteria such as ROC curves, medians, and other ambiguous methods to divide the Density of TLS into two parts, namely TLS-high and TLS-low, when using Density as the cut-off for TLS. We conducted a meta-regression to determine if different criteria would affect the predictive value of TLS. The different criteria used to divide Density can affect the predictive value of TLS(P=0.023). Further sub-analysis demonstrates that TLS has a significant correlation with OS across various criteria ( Figure S3 ).

To assess sensitivity, we employed the leave-one-out method for statistical analysis. After systematically excluding each individual study, the overall (HR) for OS, RFS, DFS, and PFS did not show any significant changes, indicating the stability and reliability of our findings ( Figure S4 ).

Next, we employed a funnel plot ( Figure S5 ), Begg’s test ( Figure S6 ), and Egger’s test ( Figure S6 ) to evaluate publication bias in the included studies. We found evidence of publication bias in OS(Begg’s test:P=0.063, Egger’s test: P=0.001)( Figures S6 A, H ). However, no publication bias was observed in RFS(Begg’s test:P=0.732, Egger’s test:P=0.430)( Figures S6 B, I ), DFS(Begg’s test:P=0.060, Egger’s test:P=0.063)( Figures S6 C, J ), and PFS(Begg’s test:P=1.000, Egger’s test: P=0.364)( Figures S6 D, K ). Further analysis of the cut-off criteria for TLS revealed publication bias in OS (Begg’s test:P=0.322, Egger’s test:P=0.033)in the “Density group”( Figures S6 F, M ). There was no publication bias in OS(Begg’s test:P=0.235, Egger’s test:P=0.064) in the “Presence group”( Figures S6 E, L ), and RFS(Begg’s test:P=0.210, Egger’s test:P=0.125) in the “Presence group”( Figures S6 F, N ). All remaining studies from the subgroups mentioned above were included in our analysis. Subsequently, we applied the trim and fill method to fill in the missing data from studies that had zero items missing. This approach ensured that our results remained robust and reliable.

We studied the relationship between TLSscore and the immune microenvironment. In the ESTIMATE algorithm, patients in the high TLSscore group showed higher immune, stromal, and ESTIMATE scores ( Figure S7 ). Single-sample gene set enrichment analysis (ssGSEA) revealed a significantly higher degree of infiltration of various immune-related cells in the TLSscore high group compared to the TLSscore low group ( Figures 5 , 6 ). Additionally, we further investigated the relationship between TLSscore and prognosis. In HCC, we found a significant improvement in overall survival (OS) associated with TLSscore high, while no significant association between TLSscore and OS was found in other gastrointestinal cancers ( Figure S8 ).