Response Rate and Survival at Key Timepoints With PD-1 Blockade vs Chemotherapy in PD-L1 Subgroups: Meta-Analysis of Metastatic NSCLC Trials

Abstract Background Expression of programmed cell death ligand 1 (PD-L1) on tumor cells with or without immune cells is widely reported in clinical trials of programmed cell death receptor 1 (PD-1) blockade in metastatic non-small cell lung cancer. Various cutpoints have been studied. Methods We performed a systematic search of MEDLINE, EMBASE, and conference proceedings up to December 2019 for randomized and nonrandomized clinical trials of anti-PD-1 or anti-PD-L1 monotherapy in metastatic non-small cell lung cancer. We retrieved data on objective response rate (ORR), 1-year and 2-year progression-free survival (PFS), and 2-year and 3-year overall survival (OS) in various PD-L1 subgroups. Results were pooled and analyzed based on different cutpoints, with nonrandomized comparisons made with pooled chemotherapy outcomes. Results A total of 9810 patients in 27 studies were included. In treatment-naïve patients, benefits with PD-1 blockade over chemotherapy were seen in ORR in patients having PD-L1 50% or greater, in 2-year OS for PD-L1 1% or greater, and in 1-year PFS, 2-year PFS, and 3-year OS for unselected patients. First-line PD-1 blockade compared with chemotherapy demonstrated higher ORR, 2-year PFS, and 3-year OS if PD-L1 was 50% or greater; lower ORR, higher 2-year PFS, and similar 3-year OS if PD-L1 was 1%-49%; and lower ORR, similar 1-year PFS, and lower 2-year OS if PD-L1 was less than 1%. In previously treated patients, PD-1 blockade demonstrated similar or superior outcomes to chemotherapy in all PD-L1 subgroups. Conclusions PD-L1 should guide the choice of PD-1 blockade vs chemotherapy in treatment-naïve patients. In previously treated patients, PD-1 blockade provides a favorable outcome profile to chemotherapy in all PD-L1 subgroups.

The most common cause of cancer death is lung cancer, with an estimated global incidence of 1.8 million and mortality of 1.6 million per year (1). Non-small cell lung cancer (NSCLC) without oncogenic driver mutations harbors a high burden of somatic mutations (2), which act as antigenic targets for host immunity. Evasion of immune destruction is a key mechanism of lung cancer development and was described as an emerging hallmark of cancer by Hanahan and Weinberg in 2011 (3). The relationship between tumor cells and host immunity usually favors tumor elimination or an equilibrium between immune destruction and tumor growth, but when tumor cells progress to immune escape, these cancer cells overcome host immunity, protect themselves with an immunosuppressive environment, and metastasize with fatal consequences (4).
Interaction between programmed cell death receptor 1 (PD-1) on T cells and programmed cell death ligand 1 (PD-L1) on tumor cells plays an important role in immune evasion. PD-1 and PD-L1 inhibitors, herein referred to as PD-1 blockade, have been established as standard of care in both treatment-naïve and previously treated patients with advanced NSCLC (5,6). Some patients experience durable tumor responses, but others derive no clinical benefit, highlighting the importance of identifying biomarkers to improve patient selection. PD-L1 expression on tumor cells with or without its expression on immune cells remains the most reported association with antitumor activity of PD-1 blockade.
PD-L1 expression is a continuous variable ranging from 0% to 100% based on the percentage of tumor cells in a tissue specimen that display partial or complete PD-L1 membrane staining of any intensity (7,8). In some assays, the percentage of tumor-infiltrating immune cells with PD-L1 staining is also measured (7,8). PD-L1 is often reported in a dichotomous fashion, where tumors are defined as either "PD-L1 high" or "PD-L1 low" based on a selected PD-L1 cutpoint. The uncertainty regarding which cutpoint should be chosen is demonstrated by the variation in cutpoints used in different studies, from 1% in some trials to 90% in others (9,10). Another important consideration is heterogeneity among different PD-L1 assays. Various pharmaceutical companies have each developed PD-1/PD-L1-inhibiting drugs with different corresponding PD-L1 assays. Results from the Phase I and II BluePrint Study, which compared the different assays, suggested that the Dako 22-8, Dako 22C3, and Ventana SP263 assays used in nivolumab, pembrolizumab, and durvalumab studies, respectively, had high concordance with each other (11). However, the Ventana SP142 assay used in atezolizumab studies had fainter staining on tumor cells and therefore produced lower scores, but in turn supplemented its scoring by including immune cell staining. The Dako 73-10 assay used in avelumab trials had stronger staining on tumor cells and therefore produced higher scores (11).
There are accumulating published data with longer followup on the relationship of PD-L1 expression with response rate, and progression-free survival (PFS) and overall survival (OS) at key time points in NSCLC patients. We conducted this metaanalysis in advanced NSCLC patients treated with PD-1 blockade, focusing on the impact of PD-L1 cutpoint choice on objective response rate (ORR), 1-year PFS, 2-year PFS, 2-year OS, and 3-year OS compared with the outcomes from chemotherapy. The aim of this meta-analysis was to identify the cutpoint at which PD-1 blockade provided superior outcomes to chemotherapy, and specifically compare PD-1 blockade with chemotherapy in the subgroups of PD-L1 less than 1%, 1%-49%, and 50% or greater.

Search Strategy and Risk of Bias
A systematic search of MEDLINE and EMBASE databases from January 1, 2010, to December 2, 2019, was performed (see the Supplementary Methods for the search strategy, available online) by the first author (J.M.) and checked by another author (J.Mi.). The search was supplemented with a hand search of abstracts and conference proceedings with the cutoff date of December 2, 2019. Studies meeting the following criteria were included: 1) phase I, II, or III human clinical trial; 2) reported results about stage IV NSCLC alone; 3) inclusion of PD-1 or PD-L1 inhibitor monotherapy arm; 4) reporting of PD-L1 subgroup analysis; 5) reporting of first-and second-line patients separately; 6) exclusion of studies limited only to patients with EGFR or ALK mutations; and 7) exclusion of studies limited to only patients with brain metastases.
Data were extracted by the first author (J.M.) and independently checked by 2 other authors (J.Mi. and A.M.), with any discrepancies resolved by consensus. In the presence of multiple publications of the same trial, the most updated data were chosen. Risk of bias was assessed by Cochrane Risk of Bias (17) for randomized trials and ROBINS-I Risk of Bias (18) for nonrandomized trials.

Statistical Analysis
Analyses were performed using Microsoft Excel v15.32. We extracted data on the primary outcomes of ORR, 1-year PFS, 2year PFS, 2-year OS, and 3-year OS in different PD-L1 subgroups of patients on PD-1 blockade monotherapy or chemotherapy. Pooled estimates and their 95% confidence intervals were calculated for the primary outcomes in each PD-L1 subgroup on each therapy. The pooled estimates were obtained as a weighted average of individual study estimates, where the study weights were determined by the inverse of the variance of the estimate in each study. This is a widely used method for pooling the results of studies and details are given in Neyeloff et al. (19). Because many studies were single arm, an analysis based on hazard ratios was not practical. Heterogeneity in the pooled estimates for each staining group was calculated by Cochran's Q test. A fixed effect method for the variance of the pooled estimates was used.
Pooled estimates of primary outcomes were then compared between the different therapies using v 2 tests. All comparisons were 2-tailed, with statistical significance defined as P less than or equal to .05. These were nonrandomized comparisons. No adjustments were made for multiple comparisons.
Cohorts of treatment-naïve and previously treated patients were analyzed separately. When extracting data for a cutpoint of interest, we used only outcomes reported in studies that were exactly consistent with the cutpoint used. For example, if data on the PD-L1 1% or greater subgroup were being pooled, outcome data reported for the PD-L1 50% or greater subgroup were not included because they may overestimate the potential benefit for the 1% or greater subgroup. When values for PFS rate or OS rate were not explicitly stated in the manuscript, these were estimated using the published Kaplan-Meier graphs. When integrating data based on the SP142 assay, which combines tumor and immune cell staining, TC0/IC0 was considered PD-L1 less than 1%, TC1/2/3 or IC1/2/3 was considered PD-L1 1% or greater, TC2/3 or IC2/3 was considered PD-L1 5% or greater, and TC3/IC3 was considered PD-L1 50% or greater.
Because chemotherapy is unlikely to be affected by PD-L1 status, comparisons of outcomes between PD-1 blockade and chemotherapy were performed between the specific PD-L1 subgroup for PD-1 blockade and all chemotherapy patients irrespective of PD-L1 status in order to optimize the power of the chemotherapy outcome estimate and allow comparisons to occur where chemotherapy arms were not reported by specific PD-L1 subgroups.
Meta-analysis was performed on all included studies. Prespecified sensitivity analyses were performed using only studies that used the Dako 22C3, Dako 28-8, and Ventana SP263 assays, which have the most comparable staining (11).

Identified Trials
A total 7255 trials were identified by the search strategy for screening ( Figure 1). A total of 170 articles or conference proceedings met the inclusion criteria, many of which were updates on the same trial. In total, 27 separate trials were included ( Table 1). Risk of bias assessment of included trials is provided in Supplementary Tables 1 and 2 (available online).
The full extracted dataset including statistical analysis can be viewed as Supplementary Material (available online for separate download).

Heterogeneity Testing
Because the number of studies being pooled for each estimate ranged from 1 to 18, and given that the fixed-effects approach yields consistent estimates of the underlying effect, the fixedeffect was the method of choice in terms of both numerical stability and interpretability given that heterogeneity was absent in the majority (>75% as calculated by Cochran's Q test) of the pooled estimates (data not shown).

Sensitivity Analyses Using Dako 22C3, Dako 28-8, or Ventana SP263 Studies Only
There was no difference in results when excluding studies that used SP142 and 73-10 assays and only including studies that used the 22C3, 28-8, and SP263 assays, which have the most comparable staining (Supplementary Figures 2-4, available online).

Discussion
This meta-analysis provides a comprehensive summary of response rate and survival at key time points for treatment with PD-1 blockade monotherapy or chemotherapy in advanced NSCLC patients with various PD-L1 expressions. An increase in publications with longer term follow-up data from trials of PD-1 blockade in advanced NSCLC has provided the foundation for this meta-analysis, which builds on the work of previous metaanalyses on this topic (63-67) by including a number of newly published phase I-III trials, updated survival data from older trials, and survival outcomes based on survival at key time points rather than median survival.
A number of key practice points are suggested by the findings of this meta-analysis. Although there is increasing evidence in the first-line setting that combination PD-1 blockade plus chemotherapy with or without the antiangiogenesis agent bevacizumab or with or without a CTLA-4 inhibitor ipilimumab is the optimal treatment for patients with PD-L1 less than 50% or those with PD-L1 50% or greater with a high burden of disease (35,(68)(69)(70)(71)(72)(73)(74)(75)(76)(77), for patients unsuitable for combination therapy or those who prefer to spare chemotherapy toxicity, the choice between PD-1 blockade monotherapy and chemotherapy remains important. This meta-analysis supports the choice of PD-1 blockade over platinum doublet chemotherapy in patients with PD-L1 50% or greater, yielding higher ORR (39.7% vs 29.0%), 2year PFS (24.6% vs 5.6%), and 3-year OS (38.3% vs 20.4%) rates. Patients with PD-L1 1%-49% are less likely to experience tumor response with PD-1 blockade than chemotherapy (16.6% vs 29.0%), but any clinical benefit is more likely to be sustained, demonstrated by higher 1-year PFS (35.0% vs 21.8%) and 2-year PFS (16.6% vs 5.6%) rates. This is an important distinction between a patient in whom urgent tumor shrinkage for symptom improvement is the primary goal of treatment compared with a patient who is asymptomatic and the chance of sustained disease control with minimal toxicity is the primary goal of treatment. The similar 2-year OS and 3-year OS rates between PD-1 blockade and chemotherapy in this group of patients, as shown in Figure 4, may be due to crossover of treatment with PD-1 blockade in later lines of therapy. In patients with PD-L1 less than 1%, most patients would benefit from chemotherapy rather than monotherapy PD-1 blockade because response rate (29.0% vs 8.3%) and 2-year OS rate (31.4% vs 16.7%) with chemotherapy are higher, and 1-year PFS rate is not any worse (21.8% vs 19.9%).
In previously treated patients, the chemotherapy tested in trials was docetaxel, which is considerably less effective than platinum doublet chemotherapy used in the first-line setting and is therefore a weaker comparator with PD-1 blockade. This meta-analysis concludes that PD-1 blockade provides equivalent or higher ORR, higher 1-and 2-year PFS rates, and higher 2and 3-year OS rates than chemotherapy regardless of PD-L1 status. Even in patients with PD-L1 less than 1%, clinicians can expect a similar ORR to chemotherapy of approximately 10% but superior 1-and 2-year PFS rates and 2-and 3-year OS rates with PD-1 blockade. Therefore, our meta-analysis supports current clinical practice guidelines that recommend the use of nivolumab and atezolizumab irrespective of PD-L1 and the use of pembrolizumab for patients with PD-L1 1% or greater, based on randomized controlled trials that demonstrate superiority of these agents over docetaxel in their respective PD-L1 populations (5,(12)(13)(14)(15)78,79).
The findings from this meta-analysis demonstrate the continuous relationship between PD-L1 expression and outcomes from PD-1 blockade. There does not seem to be a natural cutpoint above which PD-1 blockade is definitely effective and below which PD-1 blockade is ineffective. This meta-analysis therefore does not support the use of a single cutpoint to classify patients as either "PD-L1 high" or "PD-L1 low" because it can falsely separate patients close to but on either side of a cutpoint, and falsely group on the same side of a cutpoint patients who may have very different PD-L1 expression. For example, to treat a patient with PD-L1 of 40% very differently from a patient with PD-L1 of 60%, when in fact they likely would have similar responses to PD-1 blockade, can be a consequence of dichotomizing patients according to a 50% cutpoint. Throughout other branches of medicine, it has been shown that when a continuous variable is interpreted as a dichotomous variable, there can be loss of information, unproven assumption of linearity, and reduction in statistical power to determine a true relationship between the variable and its outcome (80). If more studies were to report outcomes by PD-L1 expression as a continuous variable, or at least by centiles rather than overlapping subgroups, our understanding of the true relationship between PD-L1 expression and outcomes from PD-1 blockade would be more precise. This may also create better opportunity to compare outcomes against other treatment options across the whole spectrum of the continuous PD-L1 scale without being restricted if data have only been reported by a few selected cutpoints.
This meta-analysis demonstrates the benefit of reporting survival at key time points, which can inform clinical decisions that are of direct benefit to patients. The chance of prolonged survival is an important motivator for clinicians and patients to choose immunotherapy (81), and therefore reporting the proportion of patients in the "tail of the survival curve" provides key information for clinical decision making. As more studies report survival at key time points, future meta-analyses will become more powerful.
One limitation of this meta-analysis is the inclusion of nonrandomized trials including single-arm studies to achieve an adequate sample size when comparing PD-1 blockade and chemotherapy across various PD-L1 cutpoints. Comparison of outcomes between different arms of nonrandomized trials can be affected by different baseline patient characteristics and intertrial protocol variability. Another limitation is the lack of uniformity of follow-up across trials that introduce variation in data maturity at 2-year and 3-year time points. Furthermore, publication bias is a risk, because trials showing poorer results from PD-1 blockade may be less likely to be published and updated. This meta-analysis also does not make comparison with the combinations of PD-1 blockade with chemotherapy, PD-1 blockade with cytotoxic T-lymphocyte-associated protein 4 (CTLA-4) inhibitors, or PD-1 blockade with chemotherapy and vascular endothelial growth factor (VEGF) inhibitors, which are important treatment options in the current landscape.
This study does not take into consideration a variety of other biomarkers that may also affect the efficacy of PD-1 blockade. The complexity of the immune oncology landscape is becoming increasingly evident, including the potential value of tumor mutational burden (82), immune-related gene expression signature (83,84), tumor-infiltrating lymphocyte density (85), programmed cell death ligand-2 (86), and many others.
Although this meta-analysis focuses on treatment efficacy, PD-1 blockade and chemotherapy also have different toxicity profiles, which are crucial in clinical decision making between the 2 treatment modalities. Previous meta-analyses have summarized these toxicity differences and contribute to informed treatment choices (87)(88)(89). Another area of further research includes the effect of PD-L1 expression on toxicity from PD-1blocking drugs.
Evidently, PD-L1 expression is not a perfect predictor of outcomes from PD-1 blockade in advanced NSCLC but remains an important, most studied, and validated biomarker that has clear practical implications for treatment decisions.

Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Notes
Role of the funder: Not applicable.

Data Availability
No new data were generated or analyzed in support of this research.