Abstract
Conflicting data have been reported on the prognostic value of PD-L1 protein and gene expression in breast cancer.
Experimental Design: Medline, Embase, Cochrane Library, and Web of Science Core Collection were searched, and data were extracted independently by two researchers. Outcomes included pooled PD-L1 protein positivity in tumor cells, immune cells, or both, per subtype and per antibody used, and its prognostic value for disease-free and overall survival. A pooled gene expression analysis of 39 publicly available transcriptomic datasets was also performed.
Of the initial 4,184 entries, 38 retrospective studies fulfilled the predefined inclusion criteria. The overall pooled PD-L1 protein positivity rate was 24% (95% CI, 15%–33%) in tumor cells and 33% (95% CI, 14%– 56%) in immune cells. PD-L1 protein expression in tumor cells was prognostic for shorter overall survival (HR, 1.63; 95% CI, 1.07–2.46; P = 0.02); there was significant heterogeneity (I2 = 80%, Pheterogeneity < 0.001). In addition, higher PD-L1 gene expression predicted better survival in multivariate analysis in the entire population (HR, 0.82; 95% CI, 0.74–0.90; P < 0.001 for OS) and in basal-like tumors (HR, 0.64; 95% CI, 0.52–0.80; P < 0.001 for OS; Pinteraction 0.005).
The largest to our knowledge meta-analysis on the subject informs on PD-L1 protein positivity rates and its prognostic value in breast cancer. Standardization is needed prior to routine implementation. PD-L1 gene expression is a promising prognostic factor, especially in basal-like breast cancer. Discrepant prognostic information might be related to PD-L1 gene expression in the stroma.
With the introduction of immune checkpoint inhibitors in breast cancer, there is an increased interest in PD-L1 as a predictive and prognostic marker. However, conflicting data have been reported on the prognostic role of protein expression, while few studies have investigated the prognostic value of PD-L1 gene expression. With the largest analysis of protein and transcriptomic data on PD-L1 expression in breast cancer, we highlight the heterogeneity of IHC expression according to assay and study population, as well as the discrepant prognostic information depending on the level of expression, protein, or mRNA, possibly due to its spatial expression in the stroma. PD-L1 gene expression in particular is a promising biomarker that can overcome the inherent weaknesses of evaluating PD-L1 protein expression in terms of analytic and clinical validity.
Introduction
Cancer immunotherapy through the inhibition of negative immune regulators, or checkpoints, represents a major milestone in cancer therapy. Inhibitors of Programmed Cell Death Protein 1 (PD-1; ref. 1) and its ligand (PD-L1) have been shown to significantly improve patient outcomes (2–4). Nevertheless, the fact that most patients do not derive any benefit from PD-1/PD-L1 blockade combined with the risk of serious immune-related adverse events (grade 3 and 4 in approximately 5%–10% of patients; ref. 5) and the significant upfront costs, although we so far lack sufficient long-term follow-up data for the total health–economical societal burden, underscore the need for the development of predictive biomarkers. Considering the complexity of PD-1 regulation, it is not surprising that many potential markers have been evaluated; among those, the best-studied marker is PD-L1 expression in tumor and/or immune cells, with results regarding its prognostic and predictive role being inconsistent and dependent on the tumor type, antibody, and cutoff used (6).
The routine implementation of PD-1/PD-L1 inhibition in the treatment of breast cancer has lagged behind, presumably due to its perceived lower immunogenicity (7, 8). In the only reported phase III trial, the combination of atezolizumab and nab-paclitaxel conferred a nonstatistically significant overall survival (OS) benefit compared with nab-paclitaxel alone in unselected, patients with triple-negative breast cancer (TNBC). Intriguingly, the difference in OS was statistically and clinically significant in the PD-L1–positive subgroup of patients, highlighting the potential clinical utility of PD-L1 expression as a predictive biomarker, although further validation is awaited (9, 10). On the other hand, a large number of studies investigating the prognostic implications of PD-L1 in breast cancer have been reported with conflicting results. The results of these studies have been summarized in several meta-analyses (11–15). However, the interpretation of these meta-analyses in clinical practice is limited by major methodologic drawbacks as low number of studies, inclusion of studies investigating PD-L1 expression both at the protein and mRNA levels and the inherent difficulties of evaluating a marker without robust analytic and clinical validity tested in heterogeneous populations.
In light of the aforementioned observations, we aimed to systematically investigate the incidence and the prognostic implications of the IHC PD-L1 expression in breast cancer. In addition, we performed a pooled gene expression analysis of publicly available transcriptomic datasets to investigate the correlation of mRNA expression with patient outcomes.
Materials and Methods
Search strategy and study selection
In the first part of this study, we conducted a trial-level meta-analysis of studies reporting on the prognostic value of PD-L1 protein expression. A comprehensive systematic electronic search was conducted in the following databases: Medline (Ovid), Embase, Cochrane Library (Wiley), and Web of Science Core Collection. The MeSH terms identified for searching Medline were adapted in accordance to corresponding vocabularies in Embase. Each search concept was also complemented with relevant free-text terms and these were, if appropriate, truncated and/or combined with proximity operators. Language restriction was made to English. Databases were searched from inception. The searches were performed by a librarian at the Karolinska Institutet University Library in November 2018. The search strategies are available as Supplementary Methods. We used two additional sources to ensure that all relevant studies were included: (i) the references of selected review articles on the topic were reviewed; (ii) secondary referencing by manually reviewing reference lists of potentially eligible articles.
Studies were included in our meta-analysis if they fulfilled the following criteria: studies investigating the prognostic role (measured as time-to-event outcome) of PD-L1 expression (in tumor or immune cells) in patients with early-stage breast cancer. We excluded studies investigating binary outcomes (as pathologic complete response) only. We also excluded case reports, reviews, and prior meta-analyses. Study selection was performed independently by two investigators (A. Matikas and I. Zerdes) and consensus was reached in all eligible studies.
Data extraction
Two investigators (A. Matikas and I. Zerdes) independently extracted the data on a predefined form. A third investigator (A. Valachis) compared the databases and resolved any discrepancies. The concordance rate between the two investigators was 97.1%.
The data collected from each study were: first author's last name, year of publication, country where the study was conducted, type of study (retrospective/prospective); method of PD-L1 expression, tissue used for analysis, threshold for PD-L1 expression, antibody used; positivity rate of PD-L1 in tumor cells and/or immune cells; characteristics of study cohort, follow-up time; outcome (time-to-event variables) within all patients and whenever possible within different breast cancer subtypes including both univariate and multivariate results.
Quality assessment
Two investigators (A. Matikas and I. Zerdes) independently assessed each eligible study for methodological quality using the 20-item REMARK checklist (16). Discrepancies were resolved by a third investigator (A. Valachis). The concordance rate between the two investigators considering quality assessment was 74.2%.
The REMARK checklist consists of 20 items to report in published tumor marker prognostic studies evaluating several aspects of study quality from scientific rationale and result interpretation to study design and methodology used. Each of the 20 items listed in REMARK was scored with 0 (not defined or inadequate defined or not applicable), 1 (incomplete or unclear defined), or 2 (clearly defined) for each eligible study, making a maximal score of 40.
Considering the fact that REMARK checklist has not been developed as a strict tool for assessing the study quality, we did not use the quality assessment to exclude studies from meta-analysis rather than to discuss the results in the light of the quality of eligible studies.
Statistical analysis for study-level meta-analysis
The first outcome of interest for this meta-analysis was the pooled PD-L1 positivity rate (number of cases with PD-L1–positive protein expression divided by the total number of cases, %) in tumor cells, immune cells, or both. We defined a case as PD-L1 positive when the PD-L1 protein expression by IHC was greater than the positivity threshold used in each eligible study. We used a random-effects model to produce a pooled overall PD-L1 expression positivity rate and corresponding 95% confidence interval (CI) and we then calculated the positivity rate in different subgroups of interest: breast cancer subtype based on IHC criteria [estrogen receptor (ER) positive and HER2-negative, HER2-positive and TNBC], antibody used, tissue used, threshold for positivity used. We proceeded to the calculation of pooled PD-L1 positivity rates and their 95% CI values if there were at least three studies in each subgroup. The same rule of three studies has been applied to all subanalyses in the study-level meta-analysis. χ2 statistics were used to test for differences of pooled rates among subgroups.
The second outcome of interest was the prognostic value of PD-L1 positivity based on time-to-event variables including disease-free survival (DFS), recurrence-free survival (RFS), and OS. The DFS and RFS data were merged in the meta-analysis considering the almost identical definition of these two end points (17). For the time-to-event outcomes (DFS/RFS, OS), we performed a meta-analysis first by transforming the HRs and their errors into their log counterparts, and then using the inverse variance method and transformed back into the HR scale. If time-to-event data were unavailable for direct extraction from the original publication, we extracted data according to the method described by Tierney and colleagues (18). This method allows calculation of the HR from different parameters using indirect calculation of the variance and the number of observed minus expected events. For each outcome of interest, we extracted the HRs from both univariate and multivariate analyses for each eligible study. We, then, performed meta-analyses by pooling univariate and multivariate HRs separately.
We assessed the presence of statistical heterogeneity among the studies using the Q statistics and the magnitude of heterogeneity using the I2 statistic (19). We considered a P < 0.10 or an I2 value of greater than 50% as indicative of substantial heterogeneity. When substantial heterogeneity was not observed, the pooled HR calculated based on the fixed-effects model, whereas the random-effects model was used in the presence of substantial heterogeneity.
The presence of publication bias was evaluated qualitatively using a funnel plot. All reported P values are two-sided, with significance set at P < 0.05. Statistical analyses were performed with RevMan 5.3 (Review Manager, Version 5.3; The Cochrane Collaboration, 2014) and StatsDirect (StatsDirect Ltd. UK, 2013).
Gene expression data
In the second part of this study, we conducted a pooled gene expression analysis from publicly available transcriptomic datasets. Thirty-nine datasets of gene expression profiles of more than 9,500 primary breast cancers were retrieved from public databases or authors' websites, 38 previously described by Bozovic-Spasojevic and colleagues (20) and Haibe-Kains and colleagues (21), with one additional dataset: Merck (GEO:GSE48091; ref. 22). For dataset STK, public data on histologic grade, RFS and OS (GEO:GSE1456) and in-house data on age, tumor size, and lymph node involvement were used. For dataset UPP, public clinicopathologic data (GEO: GSE3494) and in-house data on RFS and OS were used.
To ensure comparability of expression values across multiple datasets, a robust linear scaling was applied to each gene such that expression quantiles 2.5% and 97.5% were set to −1 and +1, respectively.
Each tumor was assigned to a molecular intrinsic subtype according to the PAM50 classifier (23). As no clear consensus has been established on the existence of the normal breast-like subtype, tumors of this subtype were not included in subgroup analyses.
Furthermore, as not every dataset had complete information about ER and HER2 status, ESR1 and ERBB2 status were classified on the basis of the bimodal distribution of the expression values of these two genes.
Statistical methods for gene expression analysis
Association between PD-L1 mRNA expression levels and disease-free survival [DFS, distant metastasis–free survival whenever available, RFS or progression-free interval (TCGA, as recommended by the PanCanAtlas study, https://gdc.cancer.gov/about-data/publications/pancanatlas) otherwise] and OS were analyzed. Univariate and multivariate Cox regression models with scaled expression value as continuous variable and stratification by cohort were fitted, and unadjusted and adjusted HRs and CIs were calculated. In the multivariate Cox regression model, adjustment was made for age (<50, ≥50 years), tumor size (≤2, 2–5, >5 cm), lymph node involvement (no, yes), histologic grade (G1, G2, G3) and ESR1 and ERBB2 expression status (low, high). No adjustment for ESR1 and ERBB2 status was made in models by molecular subtype. Each METABRIC study site (n = 5) was treated as separate cohort. All data analysis was done in R/Bioconductor (version 3.5.2), unless otherwise specified.
Results
Study characteristics
The flow diagram of study selection for the study-level meta-analysis is shown in Figure 1. The initial search identified 4,184 entries, or 2,745 entries following deduplication. Through exclusion by reading the title and/or abstract, 68 possibly eligible studies were retrieved as full text; 38 studies fulfilled the inclusion criteria and were included in the meta-analysis (24–61). The characteristics of eligible studies are presented in Table 1.
Flowchart of search and study selection in the meta-analysis of studies evaluating protein PD-L1 expression.
Flowchart of search and study selection in the meta-analysis of studies evaluating protein PD-L1 expression.
Characteristics of studies included in the meta-analysis of PD-L1 protein expression
| Author (reference) | Country | Type of study | Tissue for PD-L1 analysis | Antibody used | Threshold for positivity | Follow-up, median in months | Quality assessment |
|---|---|---|---|---|---|---|---|
| Acs et al. (24) | Hungary | Retrospective | Whole tissue | 28–8 | ≥1% and ≥10% | 50 | 27 |
| Adams et al. (25) | USA | Retrospective | TMA | E1L3N | >10% | NA | 25 |
| AiErken et al. (26) | China | Retrospective | TMA | E1L3N | Intensity > 1+ | 67.7 | 20 |
| Ali et al. (27) | UK | Retrospective | TMA | E1L3N | >1% | NA | 28 |
| Arias-Pulido et al. (28) | USA | Retrospective | TMA | SP142 | ≥5% | NA | 26 |
| Asano et al. (29) | Japan | Retrospective | Whole tissue | 27A2 | ≥10% | 40.8 | 30 |
| Bae et al. (30) | S. Korea | Retrospective | TMA | E1L3N | H-score ≥ 100 | 41 | 24 |
| Baptista et al. (31) | Brazil | Retrospective | TMA | 28–8 | Median | 86.2 | 30 |
| Barrett et al. (32) | USA | Retrospective | Whole tissue | 22C3 | Any staining | NA | 17 |
| Beckers et al. (33) | Australia | Retrospective | TMA | E1L3N | Membranous or cytoplasmic staining ≥1% or ≥5% | 55 | 24 |
| Botti et al. (34) | Italy | Retrospective | TMA | SP142 | ≥10% | NA | 19 |
| Brockhoff et al. (35) | Germany | Retrospective | TMA | 28–8 | Score ≥1+ | NA | 19 |
| Chen et al. (36) | China | Retrospective | Whole tissue | 28–8 | Median density | 70 | 32 |
| Choi et al. (37) | S. Korea | Retrospective | Whole tissue | EMD Millipore | ≥5% | 53 | 29 |
| Guo et al. (38) | China | Retrospective | TMA | SP142 | Score ≥1+ | 76.4 | 17 |
| He et al. (39) | USA | TMA | 28–8 | ≥1% | 45 | 30 | |
| Kim et al. (40) | S. Korea | Retrospective | TMA | E1L3N | Mean | 38.6 | 20 |
| Kitano et al. (41) | Japan | Retrospective | Whole tissue | Proscience | Any staining | 115 | 23 |
| Lee et al. (42) | S. Korea | Retrospective | Whole tissue | SP263 | H-score ≥5 | NA | 24 |
| Li M et al. (43) | China | Retrospective | Whole tissue | CST 13684 | >5% | 49 | 23 |
| Li X et al. (44) | USA | Retrospective | Whole tissue | E1L3N | Any staining | NA | 24 |
| Li Z et al. (45) | China | Retrospective | Whole tissue | 28–8 | H-score ≥100 | 64 | 29 |
| Li Y et al. (46) | USA | Retrospective | TMA | 28–8; 22C3 | ≥1% | NA | 23 |
| Mori et al. (47) | Japan | Retrospective | Whole tissue | E1L3N | ≥1% in tumor cells; ≥5% in immune cells | 68 | 27 |
| Muenst et al. (48) | Switzerland | Retrospective | TMA | 28–8 | H-score ≥100 | 65 | 33 |
| Okabe et al. (49) | Japan | Retrospective | Whole tissue | EPR1161 | H-score ≥2+ | 127.3 | 26 |
| Park et al. (50) | S. Korea | Retrospective | Whole tissue | 28–8 | H-score ≥3+ | 117.6 | 27 |
| Pelekanou 2017 et al. (51) | USA | Retrospective | Whole tissue | SP142; E1L3N | QIF score ≥500 AU | NA | 28 |
| Pelekanou 2018 et al. (52) | USA | Retrospective analysis of prospective study | Whole tissue | 22C3 | ≥1% | 36 | 26 |
| Polonia et al. (53) | Portugal | Retrospective | TMA | SP142 | ≥1% | 25 | |
| Qin et al. (54) | China | Retrospective | Whole tissue | E1L3N | >5% | 98 | 24 |
| Ren et al. (55) | China | Retrospective | TMA | SP263 | ≥25% | NA | 21 |
| Sobral-Leite et al. (56) | The Netherlands | Retrospective | Whole tissue | E1L3N | ≥1% in tumor cells; ≥5% in immune cells | NA | 29 |
| Sun et al. (57) | S. Korea | Retrospective | TMA | E1L3N; 28–8; SP142 | ≥5% in tumor cells; any staining in immune cells | NA | 22 |
| Tsang et al. (58) | China | Retrospective | TMA | Novus | Mean immunoscore | 63 (mean) | 28 |
| Wang et al. (59) | Canada | Retrospective | TMA | SP142 | H-score ≥100 | 87 | 24 |
| Wang et al. (60) | China | Retrospective | Whole tissue | ab213524 | H-score ≥100 | 66.5 | |
| Zhou et al. (61) | China | Retrospective | Whole tissue | ab213524 | Score >1+ | 45.3 | 20 |
| Author (reference) | Country | Type of study | Tissue for PD-L1 analysis | Antibody used | Threshold for positivity | Follow-up, median in months | Quality assessment |
|---|---|---|---|---|---|---|---|
| Acs et al. (24) | Hungary | Retrospective | Whole tissue | 28–8 | ≥1% and ≥10% | 50 | 27 |
| Adams et al. (25) | USA | Retrospective | TMA | E1L3N | >10% | NA | 25 |
| AiErken et al. (26) | China | Retrospective | TMA | E1L3N | Intensity > 1+ | 67.7 | 20 |
| Ali et al. (27) | UK | Retrospective | TMA | E1L3N | >1% | NA | 28 |
| Arias-Pulido et al. (28) | USA | Retrospective | TMA | SP142 | ≥5% | NA | 26 |
| Asano et al. (29) | Japan | Retrospective | Whole tissue | 27A2 | ≥10% | 40.8 | 30 |
| Bae et al. (30) | S. Korea | Retrospective | TMA | E1L3N | H-score ≥ 100 | 41 | 24 |
| Baptista et al. (31) | Brazil | Retrospective | TMA | 28–8 | Median | 86.2 | 30 |
| Barrett et al. (32) | USA | Retrospective | Whole tissue | 22C3 | Any staining | NA | 17 |
| Beckers et al. (33) | Australia | Retrospective | TMA | E1L3N | Membranous or cytoplasmic staining ≥1% or ≥5% | 55 | 24 |
| Botti et al. (34) | Italy | Retrospective | TMA | SP142 | ≥10% | NA | 19 |
| Brockhoff et al. (35) | Germany | Retrospective | TMA | 28–8 | Score ≥1+ | NA | 19 |
| Chen et al. (36) | China | Retrospective | Whole tissue | 28–8 | Median density | 70 | 32 |
| Choi et al. (37) | S. Korea | Retrospective | Whole tissue | EMD Millipore | ≥5% | 53 | 29 |
| Guo et al. (38) | China | Retrospective | TMA | SP142 | Score ≥1+ | 76.4 | 17 |
| He et al. (39) | USA | TMA | 28–8 | ≥1% | 45 | 30 | |
| Kim et al. (40) | S. Korea | Retrospective | TMA | E1L3N | Mean | 38.6 | 20 |
| Kitano et al. (41) | Japan | Retrospective | Whole tissue | Proscience | Any staining | 115 | 23 |
| Lee et al. (42) | S. Korea | Retrospective | Whole tissue | SP263 | H-score ≥5 | NA | 24 |
| Li M et al. (43) | China | Retrospective | Whole tissue | CST 13684 | >5% | 49 | 23 |
| Li X et al. (44) | USA | Retrospective | Whole tissue | E1L3N | Any staining | NA | 24 |
| Li Z et al. (45) | China | Retrospective | Whole tissue | 28–8 | H-score ≥100 | 64 | 29 |
| Li Y et al. (46) | USA | Retrospective | TMA | 28–8; 22C3 | ≥1% | NA | 23 |
| Mori et al. (47) | Japan | Retrospective | Whole tissue | E1L3N | ≥1% in tumor cells; ≥5% in immune cells | 68 | 27 |
| Muenst et al. (48) | Switzerland | Retrospective | TMA | 28–8 | H-score ≥100 | 65 | 33 |
| Okabe et al. (49) | Japan | Retrospective | Whole tissue | EPR1161 | H-score ≥2+ | 127.3 | 26 |
| Park et al. (50) | S. Korea | Retrospective | Whole tissue | 28–8 | H-score ≥3+ | 117.6 | 27 |
| Pelekanou 2017 et al. (51) | USA | Retrospective | Whole tissue | SP142; E1L3N | QIF score ≥500 AU | NA | 28 |
| Pelekanou 2018 et al. (52) | USA | Retrospective analysis of prospective study | Whole tissue | 22C3 | ≥1% | 36 | 26 |
| Polonia et al. (53) | Portugal | Retrospective | TMA | SP142 | ≥1% | 25 | |
| Qin et al. (54) | China | Retrospective | Whole tissue | E1L3N | >5% | 98 | 24 |
| Ren et al. (55) | China | Retrospective | TMA | SP263 | ≥25% | NA | 21 |
| Sobral-Leite et al. (56) | The Netherlands | Retrospective | Whole tissue | E1L3N | ≥1% in tumor cells; ≥5% in immune cells | NA | 29 |
| Sun et al. (57) | S. Korea | Retrospective | TMA | E1L3N; 28–8; SP142 | ≥5% in tumor cells; any staining in immune cells | NA | 22 |
| Tsang et al. (58) | China | Retrospective | TMA | Novus | Mean immunoscore | 63 (mean) | 28 |
| Wang et al. (59) | Canada | Retrospective | TMA | SP142 | H-score ≥100 | 87 | 24 |
| Wang et al. (60) | China | Retrospective | Whole tissue | ab213524 | H-score ≥100 | 66.5 | |
| Zhou et al. (61) | China | Retrospective | Whole tissue | ab213524 | Score >1+ | 45.3 | 20 |
NOTE: Quality assessment according to REMARK checklist, highest score is 40.
Abbreviations: PD-L1, Programmed death ligand 1; QIF, quantitative immunofluorescence; TMA, tissue microarray.
Quality of eligible studies and between-study heterogeneity
All eligible studies were retrospective. The median number of study quality score was 25 (range: 17–33) out of a maximum score of 40. Substantial between-study heterogeneity was noted among eligible studies regarding the antibody and threshold for PD-L1 positivity rate used, the follow-up period, and the study population and breast cancer subtypes.
Pooled IHC PD-L1 expression
The pooled positivity rates in overall population and in various subgroups are presented in Table 2. The overall pooled PD-L1 positivity rate was 24% (95% CI, 15%–33%; 20 studies, n = 10,404) in tumor cells, 33% (95% CI, 14%–56%) in immune cells (5 studies, n = 4,696) and 25% (95% CI, 3%–59%) in both tumor and immune cells (4 studies, n = 985). PD-L1 expression in tumor and in both tumor and immune cells was highest in TNBC. The performance of different antibodies in terms of PD-L1 expression in tumor cells was also evaluated, with Dako 28-8 clone (Agilent Technologies) yielding the highest pooled PD-L1 positivity rate, 39% (95% CI, 26%–52%, P < 0.001).
Pooled positivity rate of PD-L1 protein expression in breast cancer, as per IHC subtype, antibody, tissue, and cutoff used
| Studies (patients), n | Pooled rate, % | 95% CI | P | |
|---|---|---|---|---|
| Tumor cells | ||||
| Overall | 20 (10,404) | 24 | 15–33 | |
| According to antibody used | ||||
| 28–8 | 7 (2,251) | 39 | 26–52 | <0.001 |
| E1L3N | 6 (5,920) | 11 | 3–24 | |
| SP142 | 3 (882) | 12 | 7–17 | |
| Other | 6 (1,787) | 29 | 24–33 | |
| According to threshold used | ||||
| H-score ≥100 | 4 (1,874) | 25 | 12–40 | <0.001 |
| ≥1% | 4 (4,514) | 18 | 3–44 | |
| ≥5% | 3 (1,309) | 14 | 9–19 | |
| Mean/median | 3 (1,589) | 44 | 25–64 | |
| According to tissue used | ||||
| TMA | 9 (7,261) | 18 | 9–30 | <0.001 |
| Whole tissue | 11 (3,143) | 30 | 22–39 | |
| According to intrinsic subtypes | ||||
| ER+/HER2− | 11 (2,785) | 24 | 16–34 | <0.001 |
| HER2+ | 17 (1,279) | 30 | 22–38 | |
| TNBC | 24 (2,740) | 41 | 32–49 | |
| Immune cells | ||||
| Overall | 5 (4,696) | 33 | 13–56 | |
| According to antibody used | ||||
| 28–8 | 1 (218) | NC | NC | NC |
| E1L3N | 3 (4,355) | 24 | 3–55 | |
| SP142 | 2 (439) | NC | NC | |
| Other | 1 (120) | NC | NC | |
| According to tissue used | ||||
| TMA | 3 (4,535) | 31 | 9–59 | NC |
| Whole tissue | 2 (161) | NC | NC | |
| According to intrinsic subtypes | ||||
| ER+/HER2− | 0 | NC | ||
| HER2+ | 2 (222) | NC | NC | |
| TNBC | 10 (1,313) | 48 | 29–67 | |
| Both tumor cells and immune cells | ||||
| Overall | 4 (985) | 25 | 3–59 | |
| According to tissue used | ||||
| TMA | 1 (407) | NC | NC | NC |
| Whole tissue | 3 (578) | 31 | 5–66 | |
| According to intrinsic subtypes | ||||
| ER+/HER2− | 1 (289) | NC | NC | <0.001 |
| HER2+ | 3 (297) | 30 | 2–71 | |
| TNBC | 8 (903) | 46 | 26–67 | |
| Studies (patients), n | Pooled rate, % | 95% CI | P | |
|---|---|---|---|---|
| Tumor cells | ||||
| Overall | 20 (10,404) | 24 | 15–33 | |
| According to antibody used | ||||
| 28–8 | 7 (2,251) | 39 | 26–52 | <0.001 |
| E1L3N | 6 (5,920) | 11 | 3–24 | |
| SP142 | 3 (882) | 12 | 7–17 | |
| Other | 6 (1,787) | 29 | 24–33 | |
| According to threshold used | ||||
| H-score ≥100 | 4 (1,874) | 25 | 12–40 | <0.001 |
| ≥1% | 4 (4,514) | 18 | 3–44 | |
| ≥5% | 3 (1,309) | 14 | 9–19 | |
| Mean/median | 3 (1,589) | 44 | 25–64 | |
| According to tissue used | ||||
| TMA | 9 (7,261) | 18 | 9–30 | <0.001 |
| Whole tissue | 11 (3,143) | 30 | 22–39 | |
| According to intrinsic subtypes | ||||
| ER+/HER2− | 11 (2,785) | 24 | 16–34 | <0.001 |
| HER2+ | 17 (1,279) | 30 | 22–38 | |
| TNBC | 24 (2,740) | 41 | 32–49 | |
| Immune cells | ||||
| Overall | 5 (4,696) | 33 | 13–56 | |
| According to antibody used | ||||
| 28–8 | 1 (218) | NC | NC | NC |
| E1L3N | 3 (4,355) | 24 | 3–55 | |
| SP142 | 2 (439) | NC | NC | |
| Other | 1 (120) | NC | NC | |
| According to tissue used | ||||
| TMA | 3 (4,535) | 31 | 9–59 | NC |
| Whole tissue | 2 (161) | NC | NC | |
| According to intrinsic subtypes | ||||
| ER+/HER2− | 0 | NC | ||
| HER2+ | 2 (222) | NC | NC | |
| TNBC | 10 (1,313) | 48 | 29–67 | |
| Both tumor cells and immune cells | ||||
| Overall | 4 (985) | 25 | 3–59 | |
| According to tissue used | ||||
| TMA | 1 (407) | NC | NC | NC |
| Whole tissue | 3 (578) | 31 | 5–66 | |
| According to intrinsic subtypes | ||||
| ER+/HER2− | 1 (289) | NC | NC | <0.001 |
| HER2+ | 3 (297) | 30 | 2–71 | |
| TNBC | 8 (903) | 46 | 26–67 | |
Abbreviations: NC, not calculated; TMA, tissue microarray.
Prognostic significance of IHC PD-L1 expression
PD-L1 expression in tumor cells was prognostic for shorter DFS in the overall breast cancer population. The pooled (12 studies, n = 4,707) univariate HR for DFS was 1.36 (95% CI, 1.02%–1.83%, P < 0.04) with a significant heterogeneity as it is reflected by both I2 and Q-statistic (I2 = 73%, Pheterogeneity < 0.001). Similarly, the pooled (10 studies, n = 3,231) multivariate HR for DFS was 1.62 (95% CI, 1.14–2.33, P = 0.008; I2 = 64%, Pheterogeneity = 0.003; Supplementary Fig. S1). In addition, PD-L1 expression in tumor cells was associated with shorter OS in the overall population, in both pooled univariate (HR = 1.63; 95% CI, 1.07%–2.46%, P = 0.02; I2 = 88%, Pheterogeneity < 0.001) and multivariate (HR = 1.93; 95% CI, 1.20%–3.09, P = 0.006; I2 = 80%, Pheterogeneity <0.001) HRs (Fig. 2).
Forest plot for OS according to PD-L1 protein expression in tumor cells in total population: pooled univariate HR (A) and pooled multivariate HR (B).
Forest plot for OS according to PD-L1 protein expression in tumor cells in total population: pooled univariate HR (A) and pooled multivariate HR (B).
Further subgroup analyses are shown in Supplementary Table S1. PD-L1 expression in immune cells was associated with improved DFS in TNBC (8 studies, n = 969; HR = 0.61; 95% CI, 0.51%–0.73%, P < 0.001). The same positive association was observed for the OS endpoint as well (7 studies, n = 857; HR = 0.53, 95% CI, 0.39%–0.73%, P < 0.001). No significant heterogeneity (I2 = 17%, Pheterogeneity = 0.29 for DFS and I2 = 0, Pheterogeneity = 0.57 for OS) was observed for these pooled analyses. Because of the low number of studies, a pooled analysis of multivariate HR could not be calculated.
Publication bias
The potential presence of publication bias, namely the association between publication probability and the statistical significance of study result, was explored by visualizing asymmetry in funnel plots for each pooled analysis. There was evidence for publication bias in the pooled analyses for OS in the overall population (pooled univariate and multivariate HR) and in the Luminal, HER2-positive, and TNBC subtypes (pooled univariate HR) according to the funnel plots (Supplementary Figs. S2 and S3).
Pooled analysis of PD-L1 gene expression in relation to patient survival
In total, 39 datasets containing 9,548 patients were included for the pooled gene expression analysis. Data availability regarding PD-L1 gene expression, clinicopathologic characteristics, and information on DFS and OS is described in Fig. 3.
The prognostic value of PD-L1 gene expression per molecular subtype is shown in Fig. 4 for OS and Supplementary Fig. S4 for DFS. The corresponding univariate and multivariate analyses are presented in Supplementary Tables S2 and S3. For the DFS endpoint, data from 1,755 patients were included in multivariate analysis. Higher PD-L1 gene expression was predictive for longer DFS in multivariate analysis in both the entire population (HR = 0.69; 95% CI, 0.60–0.79; P < 0.001) and in basal-like tumors (HR = 0.53; 95% CI, 0.38%–0.74%, P < 0.001), although there was no significant interaction between PD-L1 expression and molecular subtype (Pinteraction = 0.12). Regarding the OS endpoint, data from 3,371 patients were included in multivariate analysis. Higher PD-L1 gene expression predicted better OS in multivariate analysis in the entire population (HR = 0.82; 95% CI, 0.74%–0.90%, P < 0.001) and in basal-like tumors (HR = 0.64; 95% CI, 0.52%–0.80%, P < 0.001). There was significant interaction between PD-L1 expression and molecular subtype in the OS analysis (Pinteraction = 0.005).
Forest plot for overall survival according to PD-L1 gene expression, in total population and per molecular intrinsic subtype.
Forest plot for overall survival according to PD-L1 gene expression, in total population and per molecular intrinsic subtype.
Discussion
By gathering all current evidence on the prognostic role of PD-L1 expression at the protein and mRNA level, several clinically relevant conclusions can be drawn by this meta-analysis, which, to the best of our knowledge, is the largest on the subject. In particular, the significant variation in PD-L1 expression rate depending on the study population, antibody used and positivity threshold implies an assay performance-driven heterogeneity rather than measurement of true expression. This heterogeneity is a major obstacle that needs to be overcome before PD-L1 testing can be standardized and used in daily clinical practice. In addition, prognostication based on PD-L1 expression was found to vary according to the level of expression (protein or transcriptomic) and the breast cancer subtype. Notably, no studies included in the trial-level meta-analysis reported that immune checkpoint inhibitors were used. As a result, the prognostic value of protein PD-L1 expression reported in this meta-analysis is not confounded by its putative predictive value.
The pooled positivity rates of PD-L1 expression per cell type and per breast cancer subtype provide crucial information for the design and sample size calculation of future trials. Expectedly, the pooled PD-L1 expression in TNBC, both in tumor and immune cells, was higher compared with other subtypes. These results are in agreement with several lines of evidence that support the immunogenicity of TNBC: PD-L1 mRNA expression is higher in TNBC compared with other breast cancer subtypes (62), tumor-infiltrating lymphocytes have been shown to be more abundant in TNBC and also correlate with PD-L1 expression and disease prognosis (27, 63), and clinical outcomes in TNBC have been demonstrated to be more favorable in patients with increased expression of immunomodulatory genes (64).
Interestingly, our results indicate that the performance of various antibodies differs, with Dako 28-8 clone conferring the highest pooled positivity rate compared with other antibodies. These results should be interpreted with caution, because variations in patient population, tissue used for analysis, and positivity threshold among the studies could account for the difference in PD-L1 expression. Comparative studies in breast cancer have either shown no difference between the tested antibodies (57), or higher positivity rates with Dako 28-8 clone (45). Further large-scale comparative studies across breast cancer subtypes are needed. In addition, examining whole-tissue sections yielded higher positivity rates compared with TMA cores. The underestimation of PD-L1 expression using TMA cores is explained by the significant spatial heterogeneity of PD-L1 expression (56) and supports the use of whole-tissue sections whenever feasible.
Regarding the prognostic role of PD-L1 expression, our results confirm the previously reported negative prognostic value carried by IHC PD-L1 expression in tumor cells. Considering the positive prognostic and predictive value of PD-L1 expression in immune cells in TNBC (9, 10), a single marker and its spatial expression can select populations with differential prognosis who can benefit from the addition of an effective but potentially toxic targeted therapy. In essence, our observations support the addition of PD-L1 testing as an essential marker, along with ER, PR, and HER2, in the management algorithm of metastatic breast cancer and the further categorization of TNBC in subgroups according to the applicability of immune checkpoint blockade. On the other hand, higher PD-L1 gene expression was consistently associated with improved outcomes, especially in basal-like breast cancer. Whether this phenomenon is attributed to an inherently better prognosis of immunologically active TNBC subgroups (64, 65) or to improved chemosensitivity driven by immune function (66, 67) is unclear. This positive prognostic value, similar to IHC PD-L1 expression in immune cells and in contrast with the negative prognostic value of its expression in tumor cells, suggests that PD-L1 mRNA is mainly expressed in the immune microenvironment of breast cancer therefore supporting the shift of focus to host rather than tumor expression of PD-L1 (68, 69).
This meta-analysis suffers from several limitations that need to be acknowledged. First, the meta-analysis on IHC PD-L1 expression is study-level and not individual-patient that has the advantage to minimize the between-study heterogeneity due to the ability to define exposures and outcomes consistently across studies. Second, there were significant sources of heterogeneity among eligible studies including differences in study populations, antibodies used, methods for positivity determination and type of material. The presence of heterogeneity was confirmed in most of the pooled analyses by statistical tests. As a consequence, we used random-effects model to incorporate the uncertainty due to heterogeneity among studies. Besides, we tried to assess the observed heterogeneity by subgroup analyses that revealed the necessity for standardizing the testing method. However, additional subgroup analyses of potential interest, such as prognostic value per cutoff used, could not be performed due to the limited number of studies. Finally, the presence of publication bias means that the calculated pooled effect can represent an overestimation of the true effect. Assuming that the direction of publication bias is more likely to be toward a lack of association between PD-L1 expression and prognosis, the quality of evidence from this meta-analysis should be cautiously interpreted (70). Limitations and risks for bias when conducting pooled analyses of gene expression data from public datasets are also known and well-described (71).
In conclusion, this meta-analysis highlights the variations, clinical associations and prognostic implications of PD-L1 expression at the protein and mRNA levels in breast cancer, as well as the differential prognostic information provided according to the level that PD-L1 is expressed. Efforts toward assay standardization and prospective validation are clearly needed, which could lead to refined patient management and improved outcomes thanks to the optimized use of immune checkpoint inhibition.
Disclosure of Potential Conflicts of Interest
J. Lövrot is a former employee of RaySearch Laboratories. J. Bergh reports receiving commercial research grants to his institution from Bayer, Amgen, Roche, Pfizer, Sanofi Aventis, Merck, and AstraZeneca, and reports other remuneration from UpToDate to Asklepios AB for a chapter written on prognostic factors in early breast cancer. A. Valachis reports receiving other commercial research support from Roche and reports receiving speakers bureau honoraria from Amgen, AstraZeneca, Novartis, and Roche. T. Foukakis reports receiving commercial research grants from Pfizer and Roche and reports receiving speakers bureau honoraria from Pfizer, Roche, Novartis, and UpToDate. No potential conflicts of interest were disclosed by the other authors.
Authors' Contributions
Conception and design: A. Matikas, I. Zerdes, T. Foukakis
Development of methodology: A. Valachis, T. Foukakis
Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): A. Matikas, I. Zerdes, F. Richard, C. Sotiriou, A. Valachis, T. Foukakis
Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): A. Matikas, J. Lövrot, F. Richard, J. Bergh, A. Valachis, T. Foukakis
Writing, review, and/or revision of the manuscript: A. Matikas, I. Zerdes, J. Lövrot, F. Richard, C. Sotiriou, J. Bergh, A. Valachis, T. Foukakis
Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): C. Sotiriou, J. Bergh, A. Valachis, T. Foukakis
Study supervision: A. Valachis, T. Foukakis
Acknowledgments
The authors acknowledge the contribution of Magdalena Svanberg, librarian, Karolinska Institutet University Library during the preparation of this manuscript. A. Matikas was supported by the Stockholm Region (clinical postdoctorial appointment). T. Foukakis is a recipient of the Senior Clinical Investigator Award from the Swedish Cancer Society (grant number CAN 2017/1043). J. Bergh's research group receives funding from the Stockholm region, the Swedish Cancer Society, the funds at Radiumhemmet, the Swedish Research Council, and the Knut and Alice Wallenberg fund. This study was supported by the Swedish Cancer Society (grant numbers CAN 2017/1043 and CAN 2018/846, to T. Foukakis), the Cancer Society in Stockholm (174113, to T. Foukakis), and the Stockholm Region (grant number K2017-4577, to A. Matikas).
The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
RSS Feeds