Comparative Efficacy and Safety of Programmed Death-1 Pathway Inhibitors in Advanced Gastroesophageal Cancers: A Systematic Review and Network Meta-Analysis of Phase III Clinical Trials

Simple Summary The use of checkpoint inhibitors has changed the treatment landscape for gastroesophageal cancer in the third-line setting. However, success rates in earlier treatment lines are highly variable across trials. Herein, we compare the efficacy and safety of the different anti-PD-1/PD-L1 regimens with or without chemotherapy. Abstract Background: The use of checkpoint inhibitors has changed the treatment landscape for gastroesophageal cancer in the third-line setting. However, success rates in earlier treatment lines are highly variable across trials. Herein, we compare the efficacy and safety of the different anti-PD-1/PD-L1 regimens with or without chemotherapy; Methods: We performed a network meta-analysis (NMA) of anti-PD-1/PD-L1 monotherapy or combined with chemotherapy (chemoimmunotherapy) for gastroesophageal cancers without ERBB2 overexpression; Results: The first-line NMA included four trials (N = 3817), showing that chemoimmunotherapy improved OS and PFS without significant safety difference: Nivolumab-chemotherapy, OS (HR: 0.83 [95% CI, 0.75–0.92]), PFS (HR 0.68 [95% CI, 0.57–0.81]), Pembrolizumab-chemotherapy: OS (HR 0.77 [95% CI, 0.67–0.88]), PFS (HR: 0.72 [95% CI, 0.60–0.85]. Pembrolizumab monotherapy was the safest first-line treatment, SAE (OR 0.02 [95% CI, 0.00–0.2]) but showed no survival benefit. The second-line NMA encompassed four trials (N = 2087), showing that anti-PD-1 significantly improved safety but not survival: camrelizumab, SAE (OR 0.37; [95% CI, 0.24–0.56]); nivolumab, SAE (OR 0.13, [95% CI, 0.08–0.2]) pembrolizumab, SAE (OR 0.4; [95% CI, 0.30–0.53]); Conclusions: chemoimmunotherapy improves OS and PFS in previously untreated gastroesophageal cancers. Anti-PD-1 monotherapies improve safety in refractory disease, with no significant survival benefit.


Introduction
Gastric and esophageal cancers are the third and sixth leading causes of cancer mortality worldwide, with an estimated 768,793 and 544,076 deaths in 2020, respectively [1].

Materials and Methods
We performed our study under the extension for network meta-analysis from the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) [20,21]. We created a prospective protocol and uploaded it to PROSPERO (CRD42020221822).

Eligibility Criteria
We considered eligible all randomized clinical trials comparing PD-1/PD-L1 inhibitors or anti-PD-L1, as single agents or combined with chemotherapy versus chemotherapy alone, in patients with esophageal, gastric, and gastroesophageal junction tumors, in the frontline or second-line treatments. We considered ineligible trials in phases 1 or 2 and trials that compared PD-1/PD-L1 inhibitors with other immunotherapies. When we found multiple references for the same study, we favored the latest and most complete report.

Data Sources and Extraction
We performed an extensive database search (PubMed, Embase, Cochrane Central, Web of Science, Medline, Scopus, and ClinicalTrials.gov) for entries from 1 January 2010 to 23 November 2020. We also reviewed abstracts from the American Society of Clinical Oncology and the ESMO libraries until 21 November 2020. A detailed search strategy is available in Table A1.
We uploaded titles and abstracts to Rayyan QCRI, a web-based platform for systematic review management. [22] Three authors independently performed the screening. Data from the included trials was performed by two authors, in tandem, and using a pre-piloted spreadsheet containing trial identification, baseline patient characteristics (including PD-L1 expression status), treatments, and outcomes. We resolved discrepancies by consensus. The efficacy outcomes of interest were overall survival (OS) and progression-free survival (PFS). The safety outcome of interest was the incidence of serious adverse events (SAEs), characterized as treatment-related adverse events (TRAEs) grade 3 to 5.

Risk of Bias Assessment
We used the Cochrane Collaboration's tool (version 2.0), which includes five domains (randomization process, deviation from intended interventions, missing outcome data, measurement of the outcome, and selection of reported results) and results in judgments of "low risk of bias", "some concerns", or "high risk of bias" [23]. Two authors independently applied the tool to each included trial. Any inconsistencies were solved by a discussion and between the authors.

Statistical Analysis
We performed a network meta-analysis with a frequentist approach and a randomeffects model using the package 'netmeta' for R statistical software (version 4.0.3, R Project for Statistical Computing) [24]. We used multivariate normal distribution and randomeffects models to account for between-arm correlation in multi-arm trials inside the frequentist network [25]. We generated forest plots for back-transformed network estimates. We assessed heterogeneity between and within designs using Cochran's Q statistics and quantified using I2 statistics. I2 can be used to describe the proportion of the variability in effect estimates due to heterogeneity within three thresholds 25% (low), 50% (moderate) and, 75% (high) [26,27]. We expressed OS and PFS outcomes as hazard ratios (HR) with the respective 95% confidence interval (95% CI) and SAEs as odds ratios (OR) with the respective 95% CI.

Study Selection
We found a total of 2386 unique entries. After excluding duplicates, we screened titles and abstracts for 1000 records. We assessed 149 full-text publications, including 12 trial registrations ( Figure A1). We included eight trials in the quantitative synthesis: four in the first-line setting and four in the second-line setting.

Risk of Bias
All four trials had a low risk of bias for OS. We considered CheckMate 649 a high risk of bias for PFS and SAEs, mostly related to missing outcome data. The remaining studies had a low risk of bias for PFS and SAEs (Table 2).

Study Characteristics
The studies in the second-line setting involved 2087 individuals. All trials (KEYNOTE-061, ATTRACTION-3, KEYNOTE-181, and ESCORT) compared chemotherapy with pembrolizumab, nivolumab, or camrelilzumab. The predominant tumor site was esophageal ( Table 3). All studies had subgroups according to PD-L1. All publications had OS, PFS, and SAEs data available (Table A3).

Risk of Bias
The four trials had a low risk of bias for OS and PFS. For SAE, KEYNOTE-061 raised some concerns due to missing outcome data. The remaining studies had a low risk of bias for PFS (Table 4).

Subgroup Analysis: PD-L1 Expression
Published data from PD-L1 expression subgroups across trials was not consistent. There were variable cut-off values, so it was not statistically meaningful to add those subgroups in our network meta-analysis. To evaluate PD-L1 expression as a predictor of response to ICIs, we pooled the available OS HR from chemoimmunotherapy subgroups according to PD-L1 CPS. In the first-line setting, patients that overexpress PD-L1 had better

Discussion
The present study offers valuable insight on recent advances involving the use of PD-1 inhibitors in patients with advanced gastroesophageal cancers that do not overexpress ERBB2. For previously untreated patients, chemoimmunotherapy was the best strategy. Both Nivo-Chemo and Pembro-Chemo showed significantly better OS and PFS with no significant difference in SAEs. Conversely, pembrolizumab monotherapy was markedly safer than ICI-Chemo but did not improve OS and had the worst PFS.
For patients that progressed after one line of chemotherapy, our final selection encompassed three different anti-PD-1 drugs. Camrelizumab showed the best OS, followed by nivolumab and pembrolizumab. We did not observe the same benefit for PFS. We found that camrelizumab might improve PFS, but nivolumab and pembrolizumab might worsen PFS compared to chemotherapy. Importantly, PD-1 inhibitors were significantly safer than chemotherapy as second-line treatments. Nivolumab likely has the best profile, followed by camrelizumab and pembrolizumab.
A growing body of evidence shows that drugs such as cisplatin, oxaliplatin, and paclitaxel can up-regulate PD-L1 expression in tumor and immune cells, therefore blocking the chemotherapy effectiveness but opening an opportunity to the use of PD-1/PD-L1 inhibitors [29][30][31]. Other studies point that cytotoxic therapies can turn 'cold' tumors into 'hot' tumors by making them abundantly infiltrated by CD8+ T cells and dendritic cells, making them more susceptible to ICIs [32][33][34]. The impact of these changes in the tumor microenvironment in clinical effectiveness is yet to be proven [35]. However, they help explain why chemoimmunotherapy led to better survival outcomes than pembrolizumab alone in previously untreated patients, ref. [16] while single-agent PD-1 inhibitors provided better OS benefit in patients that progressed after chemotherapy [11,12].
The use of chemoimmunotherapy for a shorter period, followed by treatment with immunotherapy only, can lead to earlier disease control with more extended survival benefits and lower SAE rates. The CheckMate 9LA has recently demonstrated the benefit of such a strategy in patients with lung cancer [31]. Ongoing phase II trials such as "Blinded for peer review" (nivolumab with or without ipilimumab) and "Blinded for peer review" (avelumab) will help to identify optimal dosing and administration schedules of immunogenic chemotherapy for gastroesophageal cancers.
Several trials in our analysis did not achieve their primary endpoints, which can be related to heterogeneity inside the cohorts [13][14][15][16]. For instance, in KEYNOTE-062, all patients had CPS ≥1, and neither Pembro alone nor Pembro-Chemo significantly improved survival. In the subset of patients with CPS ≥ 10, Pembro prolonged OS (median 17.4 months versus 10.8 months; HR 0.69; 95% CI, 0.49-0.97), however no statistical test was applied to this difference [16]. Hence, identifying which tumors will respond to immune checkpoint inhibitors is paramount. Our analysis shows that PD-L1 CPS was not a robust predictor of efficacy, as the OS benefit from chemoimmunotherapy was similar OS across subgroups, which can be related to inconsistencies in PD-L1 assessment methods and cutoff values. Hopefully, the final results from ATTRACTION-4 and the ongoing KEYNOTE-859 will help consolidate the role of PD-L1 CPS in selecting patients for chemoimmunotherapy in the first-line setting. However, there is a need for alternative biomarkers.

Limitations
Our study's first limitation comes to its nature as a network meta-analysis where we derived most of our conclusions from indirect comparisons. We used trial-level data rather than patient-level data, which could lower the power of our analysis.
Second, the trials had several differences in baseline characteristics that could affect the generalizability of the results. In the first-line, gastric and gastroesophageal adenocarcinomas were the predominant type, while in the second-line, esophageal (adenocarcinoma and squamous cell carcinoma) were the most frequent.
Third, the PFS and SAE data from CheckMate 649 included in our analysis in the first-line setting refers only to the subgroup of patients with PD-L1 CPS ≥ 5, which raises concerns for publication bias. We obtained most of the data in the first line from conference abstracts, and hopefully, further peer-reviewed publications will provide more detailed data from all patients included in each study.

Conclusions
Chemoimmunotherapy is the best first-line treatment for HER2 negative, advanced gastro-esophageal cancers. Nivo-Chemo and Pembro-Chemo improved OS and PFS similarly. Pembro did not improve survival but was significantly less toxic and should be considered as a first-line option. In the second-line setting, anti-PD-1 drugs might prolong survival, but camrelizumab was the only one to improve OS significantly. All anti-PD-1 drugs were significantly less toxic than chemotherapy for patients with refractory disease. The association of higher levels of PD-L1 expression with better outcomes remains unclear and would be better assessed in further analyses.   Web of Science ("esophageal cancer" OR "esophageal carcinoma" OR "esophageal squamous cell carcinoma" OR "gastric cancer" OR "gastric carcinoma" OR "stomach cancer" OR "gastric adenocarcinoma" OR "gastro-esophageal junction carcinoma" OR "gastro-esophageal junction cancer") AND TOPIC: ("programmed death-1" OR "pd-1" OR "pd1" OR "programmed death ligand-1" OR "pd-l1" OR "pdl1" OR "checkpoint inhibitor" OR "checkpoint blockade") AND TOPIC: ("randomized clinical trial" OR "controlled clinical trial" OR "survival rate" OR "mortality" OR "progression-free survival" OR "treatment outcome")Refined by: LANGUAGES: (ENGLISH)Timespan: All years. Indexes: SCI-EXPANDED, SSCI, A&HCI, CPCI-S, CPCI-SSH, BKCI-S, BKCI-SSH, ESCI, CCR-EXPANDED, IC.