Predictive Values of Programmed Cell Death-Ligand 1 Expression for Prognosis, Clinicopathological Factors, and Response to Programmed Cell Death-1/Programmed Cell Death-Ligand 1 Inhibitors in Patients With Gynecological Cancers: A Meta-Analysis

Background The prognostic value of programmed cell death-ligand 1 (PD-L1) in gynecological cancers has been explored previously, but the conclusion remains controversial due to limited evidence. This study aimed to conduct an updated meta-analysis to re-investigate the predictive significance of PD-L1 expression. Methods PubMed, EMBASE and Cochrane Library databases were searched. The associations between PD-L1 expression status and prognosis [overall survival (OS), progression-free survival (PFS), recurrence-free survival (RFS), cancer-specific survival (CSS) or disease-free survival (DFS)], clinical parameters [FIGO stage, lymph node metastasis (LNM), tumor size, infiltration depth, lymphovascular space invasion (LVSI) or grade] and response to anti-PD-1/PD-L1 treatment [objective response rate (ORR)] were analyzed by hazard ratios (HR) or relative risks (RR). Results Fifty-five studies were enrolled. Overall, high PD-L1 expression was not significantly associated with OS, PFS, RFS, CSS and DFS of gynecological cancers. However, subgroup analysis of studies with reported HR (HR = 1.27) and a cut-off value of 5% (HR = 2.10) suggested that high PD-L1 expression was correlated with a shorter OS of gynecological cancer patients. Further sub-subgroup analysis revealed that high PD-L1 expressed on tumor-infiltrating immune cells (TICs) predicted a favorable OS for ovarian (HR = 0.72), but a poor OS for cervical cancer (HR = 3.44). PD-L1 overexpression was also correlated with a lower OS rate in non-Asian endometrial cancer (HR = 1.60). High level of PD-L1 was only clinically correlated with a shorter PFS in Asian endometrial cancer (HR = 1.59). Furthermore, PD-L1-positivity was correlated with LNM (for overall, ovarian and endometrial cancer expressed on tumor cells), advanced FIGO stage (for overall, ovarian cancer expressed on tumor cells, endometrial cancer expressed on tumor cells and TICs), LVSI (for overall and endometrial cancer expressed on tumor cells and TICs), and increasing infiltration depth/high grade (only for endometrial cancer expressed on TICs). Patients with PD-L1-positivity may obtain more benefit from anti-PD-1/PD-L1 treatment than the negative group, showing a higher ORR (RR = 1.98), longer OS (HR = 0.34) and PFS (HR = 0.61). Conclusion Our findings suggest high PD-L1 expression may be a suitable biomarker for predicting the clinical outcomes in patients with gynecological cancers.

[hazard ratios (HR) = 2.52; 95% confidence interval (CI) =1.09 -5.83, p = 0.031] in overall or Asian patients and progression-free survival (PFS) (HR = 4.78; 95% CI = 1.77-12.91, p = 0.002) only in Asian subgroup (10). This predictive significance of positive PD-L1 expression for shorter OS (HR = 1.66) and PFS (HR = 2.17) was also demonstrated in a meta-analysis for Asian patients with ovarian cancer (12). Lu et al. reported that PD-L1 expression was significantly associated with poor differentiation (odds ratios = 2.82) and advanced International Federation of Gynecology and Obstetrics (FIGO) stage (odds ratios = 1.71) of endometrial cancer patients (11). However, there was still no meta-analysis to integrate all gynecological cancer types. More importantly, the number of included publications was relatively fewer (all < 10) in these three published meta-analyses of each gynecological cancer type (10)(11)(12). Furthermore, the clinical association of PD-L1 was not analyzed for ovarian cancer previously (12); the association of PD-L1 to anti-PD-1/PD-L1 treatment was not investigated in any type; data of tumor cells and TICs were not both collected in endometrial and cervical cancer studies (10,11) and thus their specific associations could not be performed. Hereby, the predictive performance of PD-L1 for patients with gynecological cancer remains inconclusive.
In the present study, we attempted to conduct an updated metaanalysis based on 55 published evidences to re-investigate the association of PD-L1 expression status in tumor cells and TICs with the prognosis, clinicopathological characteristics and response to anti-PD-1/PD-L1 treatment in gynecological cancer patients.

MATERIALS AND METHODS
This meta-analysis followed the guidelines of the Preferred Reporting Items for Systematic Review and Meta-analysis (PRISMA). Patient consent and ethical approval were waived because this study collected the data from published articles.

Literature Search
The online databases of the PubMed, the Cochrane Library and Embase were systematically searched up to April, 2020. The following key words were applied for searches: ("gynecological" OR "cervical" OR "ovarian" OR "endometrial") AND ("cancer" OR "carcinoma" OR "tumor") AND ("PD-L1" OR "programmed death ligand-1" OR "B7-H1" OR "CD274"). The reference lists in the retrieved papers and relevant reviews were also checked to identify additional publications.

Inclusion and Exclusion Criteria
Two reviewers independently evaluated potential articles. Studies which met the following inclusion criteria were considered eligible: 1) patients were diagnosed as any one type of gynecological cancers by pathological analyses (regardless of epithelial cancers, sarcomas or neuroendocrine tumors); 2) tumor samples for detection of PD-L1 expression were collected during primary tumor removal surgery or diagnostic biopsy before any treatment (such as neoadjuvant chemotherapy, PD-1/PD-L1 inhibitor); 3) the protein expression of PD-L1 on tumor cells or TICs of cancer tissues was determined using immunohistochemistry (IHC); 4) prognosis [OS, PFS, recurrence-free survival (RFS), cancer-specific survival (CSS) or disease-free survival (DFS)], clinicopathological parameters [FIGO stage, lymph node metastasis (LNM), tumor size, depth of infiltration, lymphovascular space invasion (LVSI), FIGO grade] and therapeutic response outcomes [objective response rate (ORR)] were compared between groups with high (positive) and low (negative) expression of PD-L1; 5) HR or relative risks (RR) as well as 95% CI values could be directly extracted, indirectly calculated using raw data or estimated from Kaplan-Meier curve; and 6) the studies were published in English and full-text. Studies were excluded if they were: 1) duplicate articles; 2) case reports, reviews, meeting abstracts, comments or letters; 3) studies evaluating the expression of PD-L1 at mRNA levels or at protein levels using other methods; 4) studies measuring the expression of PD-L1 after treatment; 5) studies having no usable data to estimate HRs and 95%CIs; 6) studies focusing on other cancers; and 7) studies written in other languages. Any disagreements were solved by discussion.

Data Extraction and Quality Assessment
Two researchers independently extracted the following data from each study: name of the first author, year of publication, country, population number, cancer type, clinicopathological features, prognostic endpoint, treatment, IHC detection area/antibody type/antibody source/IHC counting method/cut-off point for PD-L1, HRs with 95% CIs and their statistical analysis approach. Multivariable analysis results were preferentially extracted to obtain HRs and 95%CIs; otherwise, univariate analysis results were collected. The survival data in the Kaplan-Meier curves were read using a digitizing software-Engauge Digitizer 4.1. Any disputes were resolved through discussion.
The quality of included studies was assessed using the Newcastle-Ottawa Scale (NOS) (13) that consists of three key domains: selection, comparability and outcomes or exposure. Total NOS score ranged from 0 to 9. Studies with the final score > 6 were considered to have a high methodological quality.

Statistical Analysis
All data analyses were achieved with STATA 13.0 software (STATA Corporation, College Station, TX, USA). HRs with 95% CIs from each study were pooled to determine the association of PD-L1 expression with the prognostic indicators; while RRs with 95% CIs were utilized to measure the correlation of PD-L1 expression with clinicopathological factors and ORR. HR or RR > 1 indicated a poorer prognosis or higher degree of malignancy in patients with high PD-L1 expression. Association difference was analyzed using z test (p < 0.05). Heterogeneity across studies was quantified by using the Q-test and I 2 statistic. P < 0.10 and I 2 > 50% were set as the threshold for defining the studies with significant heterogeneity. A random-effect model was chosen to compute the pooled HR (or RR) for variables from studies with heterogeneity. A fixed-effect model was adopted for studies without evidence of heterogeneity. Egger's linear regression test (14) was used to detect the publication bias. If bias was seen (p < 0.05), "trim and fill" algorithm (15) was chosen for adjustment of HRs (RRs). Subgroup analysis was also carried out according to study country, sample size, cancer type, IHC detection area, antibody type, antibody source, IHC counting method, cut-off value, HR source and statistical approach to investigate possible causes of heterogeneity. Sensitivity analysis was performed via omitting any one study at a time. P-values and 95% CIs were two-sided. Figure 1 outlines the flowchart of the literature collection process. A total of 4,882 records were initially identified through searching the electronic database. After removal of 3,502 duplicate records, the titles and abstracts of 1,380 studies were read. Consequently, 1,312 articles were excluded because of they were: case reports (n = 31), meta/review (n = 47), animal studies (n = 93), studies investigating other cancers (n = 759), irrelevant topics (n = 208), without survival or other clinical outcomes (n = 172) and published in other languages (n = 2). After reviewing 68 full-text articles in detail, 16 studies were further removed since sufficient data were not provided for analysis (n = 8), IHC method was not used for detection of PD-L1 protein expression (n = 5) or the samples were collected after treatment (n = 3). Additional 3 studies were supplemented through checking the references of reviews. Finally, 55 studies were eligible for the meta-analysis . Table 1 shows the characteristics of all the included studies. The publication years ranged from 2007 to 2020 and 61.8% (34/55) of them were published within 2019 and 2020. Fourteen studies were performed in China, nine were in the USA, eight were in Japan, four in Korea, each three in Thailand, Turkey, each two in Canada, France, Germany and each one in Norway, Belgium, Brazil, Denmark, Egypt, Greece, Sweden and UK. Twenty-three studies explored the association of PD-L1 with clinical outcomes in ovarian cancer patients, 15 focused on cervical cancer and 14 investigated endometrial cancer. Ovarian and endometrial cancer patients were both enrolled in two studies, while cervical and endometrial cancer patients were both collected in one study. The prognostic endpoint was OS in 38 studies, PFS in 20 studies, RFS in 2 studies, CSS in 6 studies and DFS in 5 studies. FIGO stage (II-IV vs I or III-IV vs I-II) was compared between the groups with high and low expression of PD-L1 in 27 studies; tumor size (≥40 mm vs < 40 mm) was described in 5 studies; LNM (yes vs no) was reported in 16 studies; infiltration depth (≥ 1/2 vs <1/2) was analyzed in 7 studies; LVSI (yes vs no) was observed in 14 studies; FIGO grade was explored in 13 studies. One thing should be noted that tumor cells and TICs were both analyzed and the different IHC counting methods (cut-off points) were applied in some studies, which led to more datasets used for analysis of the prognostic and clinical significance of PD-L1 compared with the actual number of papers ( Table S1). The patients in most of these studies underwent surgery, radiotherapy and/or chemotherapy with routine drugs, while six studies specifically explored the efficacy of anti-PD-1/PD-L1 antibodies (pembrolizumab, atezolizumab, nivolumab) for the treatment of gynecological cancers (65)(66)(67)(68)(69)(70). The association of PD-L1 expression status with ORR, OS and PFS to these anti-PD-1/PD-L1 immune checkpoint inhibitors was also investigated in these six studies (65)(66)(67)(68)(69)(70). The NOS scores of all included studies were > 6, suggesting the methodological quality was high for all of them (Table S2).

Association Between Programmed Cell Death-Ligand 1 Expression and Survival
Overall Analysis in All Gynecological Cancers Fifty-one datasets (Table S1) reported the predictive values of PD-L1 for OS in all gynecological cancers. The random-effects model was chosen because of significant heterogeneity (I 2 = 71.7%, p = 0.000). The results of the meta-analysis indicated no significant association of PD-L1 expression with OS (HR = 1.13; 95% CI: 0.91 -1.39, p = 0.263). Data on PFS were extracted from 26 datasets ( Table S1). The pooled results showed that PD-L1 expression was not significantly associated with PFS (HR =

Subgroup Analysis in All Gynecological Cancers
To further investigate the possible prognostic potential of PD-L1 in gynecological cancers, the subgroup analysis was performed.

Association of Programmed Cell Death-Ligand 1 Expressions With Clinicopathological Characteristics
Overall Analysis in All Gynecological Cancers As shown in Table 4 Figure 5C).

Overall Analysis in All Gynecological Cancers
Twelve datasets reported the ORR, while OS and PFS were recorded in 5 and 7 datasets, respectively. Meta-analysis of these datasets indicated that patients with PD-L1 positive expression may get more benefit from anti-PD-1/PD-L1 antibodies than PD-L1 negative patients, showing a higher ORR (RR = 1.98; 95% CI: 1.38 -2.83, p = 0.000) ( Figure 6A

Subgroup Analysis in All Gynecological Cancers
Subgroup analysis was performed only for ORR and PFS, not OS because of small articles included. The results showed that PD-1/ PD-L1 inhibitors should be especially recommended for PD-L1-   positive ovarian patients who could gain the high ORR (n = 6: RR = 2.17; 95% CI: 1.38 -3.42, p = 0.001) and PD-L1-positive cervical patients who could obtain a longer PFS (n = 2: RR = 0.44; 95% CI: 0.29 -0.68, p = 0.000) ( Table 5). to its capability to activate the epithelial-mesenchymal transition process in a PI3K/AKT-dependent manner (76). Although previous meta-analysis studies had investigated the prognostic and clinicopathological impact of PD-L1 for cervical (10), ovarian (12) and endometrial cancer (11), the number of articles included was relatively small. Our study performed an updated meta-analysis for each gynecological cancer type by increasing the number of articles included by more than two fold. As expected, some of our results were obviously different from previous reports: our analysis showed that PD-L1 was not significantly associated with OS and PFS in any cancer type, but the study of Gu et al. reported PD-L1 overexpression was related to a poor OS in patients with cervical cancer (10); our results revealed that LNM, high FIGO stage and LVSI were more frequently observed in PD-L1-positive endometrial cancer patients compared with negative controls; while Lu et al. proved that elevated PD-L1 expression was only correlated with advanced stage, but not LVSI (11). Thus, we consider our conclusions may be more believable by analysis of larger samples. Furthermore, compared with the above mete-analyses (10,11), one innovation point in our study was to collect the PD-L1 expression on both of tumor cells and TICs, not only tumor cells. As anticipated, we obtained several new conclusions: high expression of PD-L1 on TICs was a protective factor for a poor OS in ovarian cancer patients (HR < 1), but a risk factor for unfavorable OS in cervical cancer patients, advanced stage, LVSI, high grade and increasing infiltration depth in endometrial cancer patients (HR > 1). Positive expression of PD-L1 on tumor cells was associated with a poor OS for ovarian cancer patients, LVSI for endometrial cancer patients, LNM and advanced stage for both cancer types. The anti-tumor roles of high PD-L1 on TICs for ovarian patients was also illustrated in other cancers, including colorectal (77), breast (78) and highgrade neuroendocrine carcinoma of lung (79). Its anti-cancer effects may be related with an adaptive mechanism to further activate and increase levels of cytotoxic CD8+ T cells as well as tumor-infiltrating lymphocytes (78,(80)(81)(82). Also, there was a study of non-small cell lung cancer to report that PD-L1 expression on tumor cells and TICs was associated with high levels of M2 tumor-associated macrophages and then led to a poor prognosis and an aggressive malignant phenotype, which may be one potential reason to cause the tumor-promoting effects of PD-L1 on tumor cells and TICs for gynecological cancers (83,84). In consideration of the fact that PD-L1 was highly expressed and the use of anti-PD-L1/PD-1 antibodies induced cell apoptosis and cell-cycle arrest in G0/G1 phase in gynecological cancer cells (85), increasing scholars recommended to using the PD-L1/PD-1 immune checkpoint inhibitors for the treatment of gynecological cancers in clinic (4,86). However, like other therapeutic methods, there were differences in the therapeutic efficiency among different patients (69). Thus, it is also necessary to explore biomarkers to distinguish the patients and then schedule the PD-L1/PD-1 immune checkpoint inhibitors more reasonably. Previous studies on other cancers suggested the magnitude of clinical benefit from PD-L1/PD-1 inhibitors was PD-L1-dependent (87,88). Therefore, we also investigated the associations between PD-L1 expression and ORR, OS, PFS in gynecological cancer patients. In agreement with the above studies (87-89), we also found PD-L1 patients had a significantly higher ORR (especially ovarian cancer), OS and PFS (especially cervical cancer) than PD-L1-negative patients. Although Kowanetz et al. observed that the ORR was relatively lower in patients with tumors expressing high PD-L1 levels on tumor cells than TICs (40% vs 22%) (80), our subgroup results indicated no association with tumor cells or TICs, which may be related with the small sample size. Several limitations should be acknowledged in this study. First is the retrospective nature in most of included studies. Second, the cut-off value of PD-L1 was determined by different methods in included studies, which influenced its clinical use. Third, the number of included studies to report the association of PD-L1 expression with RFS/CSS/DFS/response to anti-PD-L1/ PD-1 treatment was relatively small, which may compromise the credibility of the results and influence the subgroup analysis for each cancer type. Fourth, the estimation of HR from Kaplan-Meier curve may introduce some errors. Fifth, the restriction of articles published in other languages may lead to some negative results neglected.

CONCLUSION
Our meta-analyses ( Figure 7) indicated that positive PD-L1 detected by IHC may serve as a valuable predictor of a poor prognosis (OS, PFS), malignant clinicopathological characteristics (LNM, advanced FIGO stage and LVSI) and response efficiency to anti-PD-1/PD-L1 (ORR, OS, PFS) for patients with gynecological cancers, especially expression on tumor cells. High expressed PD-L1 on TICs may exert dual functions, including anti-cancer for ovarian cancer or oncogenic for cervical and endometrial cancers. PFS, progression free survival; ORR, overall response rate; RR, relative risk; CI, confidence interval; IHC, immunohistochemistry; TICs, tumor-infiltrating immune cells. P Z , p-value for association; P H , p-value for heterogeneity obtained by Q-test; I 2 , the degree of heterogeneity by I 2 statistic. Bold indicated the significance after analysis of two or more than two studies (p < 0.05).

DATA AVAILABILITY STATEMENT
The original contributions presented in the study are included in the article/Supplementary Material; further inquiries can be directed to the corresponding author.

AUTHOR CONTRIBUTIONS
CZ and QY conceived and designed the study, collected the data, and performed the analysis. CZ wrote the first draft of the manuscript. QY was involved in the interpretation of the analyses and revised the manuscript. All authors contributed to the article and approved the submitted version.