Comparative Efficacy and Safety of PD-1/PD-L1 Inhibitors for Patients with Solid Tumors: A Systematic Review and Bayesian Network Meta-analysis

Purpose: The clinical use of immunotherapies targeting programmed death 1 (PD-1)/programmed death ligand 1 (PD-L1) is rapid expanding, but the equivalency of these inhibitors remains unclear. We aimed to comprehensively compare the efficacy and safety of PD-1/PD-L1 inhibitors with a systematic review and Bayesian network meta-analysis Methods: We searched PubMed, Web of Knowledge, related reviews and abstracts for randomized controlled trials of five PD-1/PD-L1 inhibitors for patients with solid tumors before November 30th, 2018. We estimated summary hazard ratios (HRs) for overall survival (OS) and progression-free survival (PFS), and odds ratios (ORs) for grade 3-5 treatment-related adverse events (TrAEs) using pairwise and network meta-analysis with random-effects. This study was registered with PROSPERO (#CRD42018116624). Results: Totally, 43 reports of 35 trials comprising 21261 patients were eligible for the analysis. Nivolumab, pembrolizumab, atezolizumab and durvalumab were more effective than control treatment, and no significant differences were identified in OS and PFS between any two inhibitors. Avelumab was associated with significantly inferior OS to nivolumab (HR 1.37, 95%CrI 1.05-1.78) and pembrolizumab (HR 1.33, 95%CrI 1.02-1.73), and with inferior PFS to nivolumab (HR 1.60, 95%CrI 1.03-2.51). Compared with placebo, nivolumab had increased risk of grade 3-5 TrAEs (OR 2.35, 95%CrI 1.35-4.17). Compared with standard-of-care, nivolumab (OR 0.39, 95%CrI 0.28-0.54), pembrolizumab (OR 0.43, 95%CrI 0.30-0.60), atezolizumab (OR 0.37, 95%CrI 0.21-0.64) and avelumab (OR 0.24, 95%CrI 0.12-0.48) significantly reduced grade 3-5 TrAEs. There were not significant differences in grade 3-5 TrAEs between any two inhibitors. Conclusion: This Bayesian network meta-analysis revealed that nivolumab, pembrolizumab, atezolizumab and durvalumab yielded equivalent survival, while avelumab was associated with unfavorable survival. PD-1/PD-L1 inhibitors were comparable in the risk of TrAEs, and safer than conventional therapies.


Introduction
In the past decade, immunotherapies targeting the programmed death 1 (PD-1)/programmed death ligand 1 (PD-L1) pathway have been demonstrated to induce long-lasting survival benefits among patients with a wide spectrum of cancers, and have transformed the treatment paradigm 1,2 . To date, the U.S. Food and Drug Administration (FDA) has approved three PD-1 inhibitors (nivolumab, Ivyspring International Publisher pembrolizumab, and cemiplimab) and three PD-L1 inhibitors (atezolimumab, durvalumab and avelumab) for the treatment of solid tumors in different clinical settings, including non-small-cell lung cancer (NSCLC), melanoma, urothelial carcinoma, head and neck squamous cell carcinoma (HNSCC) and merkel cell carcinoma (MCC). Moreover, nivolumab and pembrolizumab are recommended as the treatment of choice in the first-line setting for melanoma and NSCLC by the National Comprehensive Cancer Network (NCCN) Clinical Practice Guidelines in Oncology (https://www.nccn.org/), as phase 3 randomized trials have proven the superiority over standard of care (SoC). They also provide clinically significant survival advantage among patients with stage III melanoma after complete resection 3,4 , and are category 1 options for adjuvant therapy. Despite the remarkable clinical success of anti-PD-1/PD-L1 immunotherapy, it is unclear whether these inhibitors have the same effects on restoring anti-tumor immunity, resulting in the difficulty in choosing the best inhibitor in the clinic. In the aspect of tumor immune biology, PD-1 inhibits signaling downstream of the T cell receptor to regulate immune cell activity within tissue and tumors, when engaged by its two ligands, PD-L1 and PD-L2 5 . In addition to interacting with PD-1, PD-L1 also interacts with CD80 expressed on T cells and mediates an inhibitory signal 6 . What's more, PD-L1 could function as a receptor to mediate tumor cell-intrinsic signals to affect cell growth in the absence of immunity 7,8 . PD-1 inhibitors could block PD-1 from interacting with both PD-L1 and PD-L2, but not the PD-L1/CD80 interaction. By contrast, PD-L1 inhibitors could block the PD-L1/PD-1 and PD-L1/CD80 interactions, but leave PD-L2/PD-1 pathway intact. Additionally, PD-L1 inhibitors could also directly interfere the immune-independent cell growth mediated by PD-L1. Lastly, PD-L1 antibody has the potential to induce antibody-dependent cellular cytotoxicity when binding to the surface of tumor cells, contributing to its clinical activity 9 .
In addition to the differences of the molecular biological functions of PD-1/PD-L1 inhibitors, differences of their performance in clinical trials were noted. The optimal efficacy of pembrolizumab was observed in treated NSCLC patients with PD-L1 expression in at least 50% of tumor cells, revealing the predictive value of PD-L1 expression as a biomarker 10,11 . Nevertheless, nivolumab 12 or atezolizumab 13 treatment led to a clinically relevant improvement of survival versus docetaxel in treated NSCLC patients, regardless of PD-L1 expression level. Besides, pembrolizumab monotherapy significantly improved the survival over chemotherapy in treatment-naive advanced NSCLC patients with PD-L1 strong positive expression 14 , while nivolumab monotherapy failed to extend survival in the first-line trial 15 . Collectively, these findings bright up the question whether PD-1/PD-L1 inhibitors have equivalent efficacy and safety.
To the best of our knowledge, there have been not randomized trials published or in progress which "head-to-head" compare these PD-1/PD-L1 inhibitors so far. The lack of robust data evaluating the efficacy and toxicity of these inhibitors is a major dilemma of the clinic practice. Bayesian network meta-analysis of present evidence uses a common comparator to achieve indirect comparison when a head-to-head clinical trial is absent, and combines direct and indirect comparisons to simultaneously compare multiple interventions 16 . Therefore, we performed a systematic review and network meta-analysis of all randomized trials to compare the efficacy and safety of the FDA-approved PD-1/PD-L1 inhibitors for patients with solid tumors.

Study eligibility and selection
We conducted the systematic review in line with the recommendations of the Cochrane Collaboration the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. The protocol was prospectively registered with the PROSPERO database (registration ID: CRD4201 8116624). All randomized controlled trials that comparing the overall survival (OS) or progression-free survival (PFS) of PD-1/PD-L1 inhibitors monotherapy verse (vs) standard of care (SoC) or placebo, PD-1/PD-L1 inhibitors combined with SoC vs SoC in adult patients with solid tumors were eligible for analysis. The approval of cemiplimab in September 2018 was based on two single-arm trials 17 , and no randomized trials had been reported, so it was not included in present study.
Potentially relevant studies were identified from previous meta-analyses [18][19][20] , and a search of PubMed and Web of Knowledge (through November 30 th, 2018), with no limitations on publication year or language restrictions. Abstracts from the major conference proceedings of the American Society of Clinical Oncology, the European Society for Medical Oncology, and the World Conference for Lung Cancer were also searched for unpublished studies. We used search terms "carcinoma", "cancer", "trial", "checkpoint inhibitors", "PD-1", "PD-L1" combined with PD-1/PD-L1 inhibitors' names (nivolumab, pembrolizumab, atezolizumab, durvalumab and avelumab). The search strategies were detailed in the supplementary materials. We combined the search results in a bibliographic management tool (EndNote), and eliminated duplicates. After preliminary screening of titles and abstracts, three independent authors (Q-YH, Y-ZZ and Z-DG) assessed the full text for final selection. We also reviewed the references of articles included in the final inclusion.

Data extraction and quality assessment
Data extraction was carried out using a predefined electronic database. Both OS and PFS were chosen as the outcomes for efficacy, and the HRs and their 95% confidence interval (95%CI) were our preferred outcome measure, because HRs provide time-to-event information and account for censoring. Grade 3-5 treatment-related adverse events (TrAEs) were chosen as the outcomes for safety, and number of patients in safety analysis and grade 3-5 TrAEs of each treatment arm were recorded. Multiple source reports derived from one trial were all included in selection process, and the most updated and complete data were extracted for the analysis. Besides, we also extracted other clinicopathological characteristics of each study, including first author, year of publication or presentation, study name, cancer type, trial phase, line of therapy, treatment regimen, sample size and median of follow-up.
We employed the Jadad scale to quantitatively evaluate the quality of studies. Study selection, data extraction and quality assessment were performed by two of three authors (Q-YH, Y-ZZ and Z-DG). Any discrepancies were resolved by discussion involving all three to achieve consensus.

Data synthesis and statistical analysis
We estimated summary HRs for time-to-event outcomes, and odds ratios (ORs) for dichotomous outcomes using traditional pairwise and network meta-analysis. Pooled HRs of multiple doses of one drug in one trial were combined by meta-analysis. As the included trials are likely different in many aspects, heterogeneity is generally expected in a meta-analysis. Therefore, the study effect sizes were synthesized using random-effects models to account for heterogeneity.
Bayesian network meta-analysis was performed with the "rjags" package in R (version 3.5.1), which used Markov Chain Monte Carlo (MCMC) algorithm with three chain to generate posterior distribution of model parameters. Each chain had a run of 200,000 updates after a 50,000-run burn-in period with thinning rate 2 for reducing sample autocorrelations. We performed pairwise meta-analysis to synthesize studies comparing the same pair of treatments with Stata 12 (StataCorp, College Station, TX, USA). Excellent consistency, represented by the agreement between indirect and direct comparisons, is the key to robust results. Consistency was evaluated by comparing the pooled HRs from the network meta-analysis with corresponding HRs from pairwise meta-analysis. P values (less than 0.05) and 95% CIs/CrIs were used to assess statistical significance. All statistical analyses were two-sided.

Study selection and characteristics
We identified 3583 citations for review of title and abstract by the search. As shown in Figure 1, the screening process totally included 43 publications reporting 35 trials of 21261 patients, evaluating five PD-1/PD-L1 inhibitors (23 trials of PD-1 inhibitors: 11 of nivolumab and 12 of pembrolizumab; 12 trials of PD-L1 inhibitors: 7 of atezolizumab, 2 of durvalumab and 3 of avelumab). Eight trials were conducted in melanoma, 16 in lung cancer (8 in NSCLC, 5 in non-squamous cell NSCLC, 2 in squamous cell NSCLC and 1 in small-cell lung cancer), 3 in HNSCC, 4 in urinary system cancer, 3 in gastric or gastro-esophageal junction cancer and 1 in triple-negative breast cancer. There were 4 phase 2 trials, 29 phase 3 trials, and 1 phase 2/3 trial. Fourteen trials were done in first-line setting, 17 in second-or subsequent-line setting, 1 in first-or second-line setting and 3 in adjuvant setting. The median score was 3 (range 2-5), with 31 (88.6%) reports receiving high-quality score (Jadad score of [3][4][5], and the remaining four reports with low-quality score (Jadad score of 1-2) were all meeting abstracts without detailed reports at this moment (Supplementary Table  S1). Figure 2A showed the network of drug-based comparison, which was designed for multiple comparison of the five inhibitors. Figure 2B showed the network of category-based comparison, which established the comparison of PD-1 inhibitors (nivolumab and pembrolizumab) and PD-L1 inhibitors (atezolizumab, durvalumab and avelumab).
We performed subgroup analyses according to the line of therapy (Supplementary Figure S1), type of control treatment (Supplementary Figure S2) and cancer type (Supplementary Figure S3). Significant differences in OS and PFS were not noted between any two drugs of nivolumab, pembrolizumab, atezolizumab and durvalumab, except that atezolizumab had significantly unfavorable PFS than nivolumab (HR 1.53, 95%CrI 1.03-2.28) in the placebo-controlled subgroup. Poorer OS and PFS than nivolumab and pembrolizumab were also noted with avelumab, although statistical significance was not reached in some subgroups. It was not applicable to pool the risk of death or progression for some subgroups, as only one trial was available in these subgroups, such as urinary system cancer subgroup and HNSCC subgroups.
CheckMate 067, CheckMate 069, KEYNOTE-006, and CheckMate 238 trials used ipilimumab, an antibody against CTLA4, as the control regimen. Sensitivity analysis by excluding these trials derived consistent results with the original analysis in terms of survival and toxicity (Supplementary Figure S4).

Category-based network meta-analysis
We further classified the drugs as PD-1 inhibitors or PD-L1 inhibitors according to the treatment target. The network meta-analysis ( Figure  5A Figure S7) were also performed. We did not find significant difference in OS or PFS between PD-1 inhibitors and PD-L1 inhibitors in all subgroups. Previous analysis indicated that the efficacy of avelumab was inferior to nivolumab and pembrolizumab, therefore, we excluded avelumab and re-analyzed the data. Re-analyasis excluding trials of avelumab did not show any significant differences in OS (HR 1.11, 95%CrI 0.95-1.29) or PFS (HR 1.08, 95%CrI 0.84-1.40) between PD-1 inhibitors and PD-L1 inhibitors.

Discussion
The clinical use of anti-PD-1/PD-L1 immunotherapy is rapidly expanding, and more drugs are striving to enter the game 57 , however, little comparative evidence of the inhibitors is currently available. Our Bayesian network meta-analysis compares five FDA-approved PD-1/PD-L1 inhibitors (cemiplimab was excluded for unavailable randomized trials), and analyzes both survival and toxicity of the inhibitors. We found that all PD-1/PD-L1 inhibitors, except avelumab, were more efficacious than control treatment in patients with solid tumors. Significant inferior survival was observed among patients treated with avelumab than those with nivolumab and pembrolizumab, while the other four inhibitors treatment yielded equivalent survival. These findings were further strengthened by their consistency across all analyzed subgroups, as the large sample size allowed thorough subgroup analyses. Analysis of toxicity demonstrated no significant differences in the risk of grade 3-5 TrAEs between any two inhibitors. It was noteworthy that all five inhibitors were generally associated with lower frequency of TrAEs when compared with SoC, making them more favorable options in the clinic.
Avelumab was approved by FDA for metastatic MCC and urothelial carcinoma based on its excellent performance in two single-arm trials. Among 88 patients with treated MCC who were enrolled for avelumab treatment, 32% experienced complete or partial remission of their tumors, and responses were ongoing in 82% of responding patients at the time of analysis 58 . FDA granted avelumab an accelerated approval and orphan drug designation to fill an unmet medical need for MCC. The approval for urothelial carcinoma was based on data of 242 patients, of whom the overall response rate (ORR) was 13.3% among those who were followed for at least 13 weeks, and 16.1% among who were followed for at least 6 months 59 . Nevertheless, avelumab treatment failed to reach the study endpoint of OS or PFS improvement among patients with NSCLC 37 and gastric or gastro-esophageal junction cancer 55 in the subsequent-line setting. Considering these results, our findings are reasonable that the survival with avelumab treatment was not superior to that with chemotherapy, and inferior to that with nivolumab and pembrolizumab treatment. These results indicate that PD-1/PD-L1 inhibitors are not equally efficacious, and the efficacy of an inhibitor may vary with cancer type. These inhibitors should not be simply regarded as one entity, although they principally act through blocking PD-1/PD-L1 pathway. More attention should be paid to this in clinical practice and design of clinical studies.
To the best of our knowledge, the present study has been the most comprehensive and updated meta-analysis so far, and is the first one to directly reveal the disparity in survival of PD-1/PD-L1 inhibitors. Two indirect comparisons by Passiglia et al 60     In addition, this work is also strengthen by the quality and quantity of studies identified and used in the pooled analysis. Data are obtained from 35 published and unpublished randomized controlled trials, of which 88.6% achieve high-quality Jadad scores. OS and PFS are used to measure the efficacy outcomes, and grade 3-5 TrAEs to the safety outcomes. These endpoints are well-defined in clinical trials. The large number of included trials without a limitation of cancer type is a crucial point. Recently-approved drugs, such as durvalumab and avelumb, have a small number of randomized trials at this moment. Our high quality and large quantity of data lead to good statistical power, and make the results reliable. Finally, all pooled HRs and their CrIs of network meta-analysis are close to the corresponding results of pairwise meta-analysis, supporting that there is no significant inconsistency between direct and indirect comparisons.
Our work has some limitations. First of all, the present analysis included all solid tumors with different control treatment regimens, and this would introduce heterogeneity to the results. To address this issue, we performed detailed subgroup analyses, and generate same results. Additionally, two trials of PD-1 inhibitors as adjuvant treatment, CheckMate 238 4 and KEYNOTE-054 3 , only reported recurrence-free survival results, which are used as surrogate endpoint of PFS in our analysis. The OS data are immature, thus unavailable for the synthesis of. Lastly, this meta-analysis was conducted with summary statistics rather than individual participant data, so it could not be adjusted for some potential covariates. In summary, our large-scale network meta-analysis based on Bayesian approach revealed the disparity in the efficacy of PD-1/PD-L1 inhibitors. The survival of treatment with nivolumab, pembrolizumab, atezolizumab and durvalumab were equivalent in solid tumors, while avelumab was associated with unfarvorable survival. PD-1/PD-L1 inhibitors were comparable in the risk of TrAEs, and safer than conventional therapies. These results could assist in the decision-making of clinicians and their patients, and provide more information for the future clinical studies.