Prognostic Role of PD-L1 Expression in Invasive Breast Cancer: A Systematic Review and Meta-Analysis

Simple Summary The role of PD-L1 expression in breast cancer remains controversial. Therefore, we performed a systematic review and meta-analysis to assess the association of PD-L1 expression with clinicopathological variables, overall survival (OS), and disease-free survival (DFS) in invasive breast cancer. PD-L1 expression was associated with age ≥ 50 years, lymph node status-negative, progesterone receptor-negative, Ki67 ≥ 20%, and human epidermal growth factor receptor 2 (HER2)-negative. PD-L1 positivity was associated with worse OS; however, there was no significant improvement in DFS. PD-L1 positivity was significantly associated with the clinicopathological characteristics of favorable and unfavorable prognoses. However, the final clinical outcome was associated with lower OS and had no significant association with DFS. Abstract Programmed death ligand 1 (PD-L1) has been investigated in various types of cancer; however, the role of PD-L1 expression in breast cancer remains controversial. We performed a systematic review and meta-analysis to assess the association of PD-L1 expression with clinicopathological variables, overall survival (OS), and disease-free survival (DFS) in invasive breast cancer. A total of 965 articles were included from CINAHL, Embase, PubMed, and Scopus databases. Of these, 22 studies encompassing 6468 cases of invasive breast cancer were included in the systematic review, and 15 articles were included in the meta-analysis. PD-L1 expression was associated with age ≥ 50 years, lymph node status-negative, progesterone receptor-negative, Ki67 ≥ 20%, and human epidermal growth factor receptor 2 (HER2)-negative. PD-L1 positivity was associated with worse OS (hazard ratio, HR, 2.39; 95% confidence interval, CI, 1.26–3.52; p =< 0.000); however, there was no significant improvement in DFS (HR 0.17; 95% CI −0.12–0.46; p =< 0.252). PD-L1 positivity was significantly associated with the clinicopathological characteristics of favorable and unfavorable prognoses. However, the final clinical outcome was associated with lower OS and had no significant association with DFS.


Introduction
The development of immunotherapy provides a new mechanism of action within systemic cancer therapy, as opposed to conventional treatments that lack tumor selectivity and cause adverse side effects [1,2]. These therapies use monoclonal antibodies against Table 1. Database search strategy.

Databases Search Strategy
Medline/PubMed

Eligibility Criteria
The inclusion criteria were (1) observational or interventional studies on PD-L1 expression in IBC; (2) studies evaluating the prognostic ability of PD-L1 expression by immunohistochemistry; (3) studies without language restriction; and (4) articles published between 1 January 2018, and 28 January 2021. The exclusion criteria included theses, dissertations, case studies, animal studies, reviews, editorials, letters to the editor, duplicate studies, studies with specific populations (e.g., pregnant or lactating women), studies that evaluated rare histological types, and studies restricted to HER2-positive or triple-negative molecular subtypes. The full texts of the articles were requested from the authors [11]. There were no restrictions based on the treatment received for BC.
The search was limited to articles published in the period between August 2018 and January 2021 due to the existence of other systematic reviews based on prior data [1,9,11]. This study focused on discussing the latest evidence on this topic.

Study Selection
Titles and abstracts were screened by two researchers (M.B.C. and C.R.M.) using Rayyan software. The articles were read in full and selected according to the inclusion and exclusion criteria. Disagreements on the quality of evidence were discussed amongst the research team.
The following data were extracted: study authors and year of publication, experimental design, country, number of patients, age, sample type, data evaluation methods, and clinical outcomes (clinicopathological characteristics and survival-OS and DFS). The clinicopathological variables included age, tumor size, lymph node status, estrogen receptor (ER) status, progesterone receptor (PR) status, HER2 status, Ki-67 index, and molecular subtypes (luminal A and B, HER2-negative, and triple-negative).

Risk of Bias and Analysis of the Quality of Evidence
Risk of bias (RoB) was assessed in primary-level studies using the Quality in Prognosis Studies (QUIPS) tool, supported by Cochrane Prognosis Methods Group for prognosis studies [12,13]. QUIPS considers the following domains: (1) study participation, (2) study attrition, (3) prognostic factor measurement, (4) outcome measurement, (5) study confounding, and (6) statistical analysis and reporting.
The quality of the scientific evidence was evaluated using the Grading of Recommendations, Assessment, Development, and Evaluation (GRADE) online software (https: //gdt.gradepro.org/app/#/, accessed on 26 September 2021) [14,15] and was classified as high, moderate, low, or very low [16].

Training of the Reviewers
The authors who participated in the eligibility assessments were trained regarding the study inclusion/exclusion criteria and completed a practice eligibility assessment on 50 test abstracts before starting to code articles. Moreover, the authors were also trained in performing risk of bias instruments on five articles not included in the study as well as standardized analyses using Mendeley and Rayyan software [17].

Statistical Analysis
A meta-analysis was conducted using the random-effects model for coded and stratified data on PD-L1 expression in the following cell types: tumor cells (TCs), immune cells (ICs), and (c) both tumor cells and immune cells (TCICs). The proportion of PD-L1 expression was determined in TCs and ICs according to clinicopathological variables, and the hazard ratios for OS and DFS were calculated. Proportion rates and hazard ratios with 95% confidence intervals (CIs) are shown as forest plots. To calculate the proportion, we use the command metaprop, grouping proportions which are specific to binomial data, allowing computation of exact binomial and test-based CIs. The degree of heterogeneity (I 2 ) between the studies was calculated. I 2 < 25%, I 2 = 25-50%, and I 2 > 50% indicated low, moderate, and high heterogeneity, respectively [18]. Publication bias was assessed using Egger's test and funnel plots. All analyses were performed using STATA software (version 16.0; StataCorp, College Station, TX, USA).

Identification of Studies
A total of 965 articles were identified through database searches, and five additional studies were identified through reference lists. After removing duplicates, 662 articles were selected for reading the titles and abstracts. The decisions of the first researcher (M.B.C.) and second researcher (C.R.M.) were compared, and a Cohen kappa statistic indicated high concordance between them (93.04%; adjusted kappa, 0.80) [19]. Seventy articles were selected for full-text reading. Of these, 22 articles met the eligibility criteria and were included in the systematic review, and 15 articles were included in the meta-analysis. The flowchart of the study selection process according to the PRISMA guidelines is shown in Figure 1.

Study Characteristics
A total of 22 articles were included in this study . Most of th retrospective, and the follow-up interval varied from 3 months to 15 years. BC cases from Italy, Greece, Japan, Korea, Germany, Egypt, the Netherlan States, Sweden, and China were included in the analysis. Information on th anti-PD-L1 clones, immunohistochemical analysis criteria, and cell types Table 2. No patients received immunotherapy in the included studies.

Study Characteristics
A total of 22 articles were included in this study . Most of the studies were retrospective, and the follow-up interval varied from 3 months to 15 years. A total of 6468 BC cases from Italy, Greece, Japan, Korea, Germany, Egypt, the Netherlands, the United States, Sweden, and China were included in the analysis. Information on the sample type, anti-PD-L1 clones, immunohistochemical analysis criteria, and cell types is presented in Table 2. No patients received immunotherapy in the included studies.

HER2 Status
The frequency of PD-L1 expression in TCs was significantly higher in HER2-negativ

PD-L1 Expression and Clinicopathological Characteristics
The proportion of PD-L1 expression in TCs and ICs was determined according to the following clinicopathological variables: age, tumor size, lymph node status, hormone receptor (ER, PR, HER2) status, Ki-67 index, and molecular subtypes (luminal A and B, HER2-positive, and triple-negative). The results were extracted from the meta-analysis graphs and are summarized in Table 4. All meta-analysis graphs that analyzed the expression of PD-L1 in TCs and in ICs according to clinical-pathological characteristics are available in Figure S1.

Quality Assessment and Risk of Bias
The quality of scientific evidence was evaluated with the GRADE quality assessmen tool, and the risk of bias was calculated using the Cochran Collaboration Risk of Bias To for non-randomized studies. The analysis of GRADE scores indicated that two studies ha a high quality of evidence [30,38], 15 studies presented moderate quality, and four studi had low quality of evidence [21,34,36,37] (Table 1).
The risk of bias RoB scores are shown in Figure 5. The risk of bias was moderate t low in most studies. The low quality of evidence and high risk of bias were due to sma sample size and conflicting data.

Quality Assessment and Risk of Bias
The quality of scientific evidence was evaluated with the GRADE quality assessment tool, and the risk of bias was calculated using the Cochran Collaboration Risk of Bias Tool for non-randomized studies. The analysis of GRADE scores indicated that two studies had a high quality of evidence [30,38], 15 studies presented moderate quality, and four studies had low quality of evidence [21,34,36,37] (Table 1).
The risk of bias RoB scores are shown in Figure 5. The risk of bias was moderate to low in most studies. The low quality of evidence and high risk of bias were due to small sample size and conflicting data.

Publication Bias
Publication bias was evaluated using a funnel plot ( Figure 6). The 15 studies included in the analysis had little publication bias. This finding was confirmed with the results from Egger's test (p = 0.810).

Publication Bias
Publication bias was evaluated using a funnel plot ( Figure 6). The 15 studies included in the analysis had little publication bias. This finding was confirmed with the results from Egger's test (p = 0.810). Figure 5. Quality plot graphically representing the risk of bias (RoB) analysis. The most r sources of bias were assessed in primary-level studies using the Quality in Prognosis (QUIPS) tool.

Publication Bias
Publication bias was evaluated using a funnel plot ( Figure 6). The 15 studies in in the analysis had little publication bias. This finding was confirmed with the result Egger's test (p = 0.810).

Discussion
We identified and analyzed studies that investigated the clinicopathological features and prognostic ability of PD-L1 expression in IBC. Data from 22 studies involving 6468 BC cases were evaluated. As expected, we observed great heterogeneity between studies in relation to BC characteristics, pathologic material analyzed (TMA or full section), anti-PD-L1 clone used, determination criteria, and follow-up time. Thus, the data reported in this study will contribute to the understanding of the prognostic role of PD-L1 and its evaluation by immunohistochemistry (IHC).
The difference between the clones used and the material analyzed plays a crucial role in the rate of PD-L1 positivity, both in BCs and other tumors [42,43]. Several studies have evaluated the inter-assay variability between the different tests available for the analysis of PD-L1 expression, most with moderate correlation [22,42,44]. The SP142 assay, which predicts the response to atezolizumab in triple-negative BC, shows high interobserver agreement [45]. However, almost 30% of tumors considered PD-L1-positive by SP263 or 22C3 tests are negative by SP142 [42,45]. Regarding the analyzed pathological material, it should be noted that many of the included studies performed their analyses by TMA. Nevertheless, up to half of PD-L1 scores evaluated by TMA can be false negatives compared to whole slide evaluations [33].
The proportion of PD-L1 expression was higher in TCICs (37%), followed by ICs (30%) and TCs (19%). Another systematic review found that the frequency of PD-L1 expression was 25.8%, although the cell type studied was not reported [11]. Considering that the first indications of immunotherapy for BC may prioritize this population, defining the real proportion of PD-L1 is the first step towards the development of clinical protocols and public policies for access to medications.
Few studies have evaluated PD-L1 expression in TCICs according to their clinicopathological features. Only one study analyzed the proportion of PD-L1 expression in TCICs [22]. This study demonstrated that PD-L1 expression was significantly more prevalent in ER-positive (65.7% vs. 34.3%, p = 0.003), PR-negative (57.1% vs. 42.9%, p ≤ 0.000), and HER2-negative tumors (82.9% vs. 17.1%, p = 0.018) [22]. In clinical practice, the expression of TCICs has been commonly described using the combined positive score (CPS), which is the number of PD-L1 staining cells (TCs, lymphocytes, and macrophages) divided by the total number of viable TCs, multiplied by 100 [43]. In the KEYNOTE 355 study, which included cases of metastatic BC with triple-negative tumors, the addition of pembrolizumab was observed to significantly improve PFS compared with chemotherapy alone in patients with CPS ≥ 10 [46]. However, in a prognostic context, controversies remain about the most appropriate cutoff. In the present review, the studies that analyzed the expression of PD-L1 in TCICs used the proportional percentage and the 1% cutoff.
PD-L1 expression was higher in TCs and ICs in women older than 50 years, with no significant differences between the cell types. Furthermore, the frequency of PD-L1 expression was higher in ICs in patients with node-negative status. In triple-negative BC, PD-L1 positive tumors have more immunogenic characteristics, including elevated tumor infiltrating lymphocyte (TIL) and CD8 counts, enrichment of the immunogenic genomic subtype, and elevated immunogenic gene signatures at the gene expression level [47]. However, it is not clear whether these characteristics could determine the occurrence of tumors at older ages or early stages. It should be noted that the SP142 assay detects more ICs and fewer TCs compared to the other assays, which can generate conflicting results depending on the PD-L1 assay utilized [47]. Furthermore, the interobserver agreement for PD-L1 expression in ICs is inferior to TCs in various types of tumors regardless of the type of assay used, which could also contribute to the divergence of prognostic factors [3,44].
The proportion of PD-L1 expression was higher in TCs in PR-negative cases and in cases of TCs and ICs with a Ki67 index ≥20% and HER2-negative status. Another review found that PD-L1 upregulation was associated with high-risk prognostic factors, including high histological grade (p = 0.000), ER negativity (p = 0.000), PR negativity (p = 0.000), HER2 positivity (p = 0.001), and aggressive molecular subtypes (HER2-positive and triple-negative; p = 0.000) [11]. In parallel, a meta-analysis that included different types of epithelial-originated cancers observed an 81% increased mortality risk in a group of tumors with positive PD-L1 expression. However, the prognostic impact of PD-L1 status is more evident when stricter criteria for positive PD-L1 expression are applied [3], which reinforces the need to standardize cutoff values for each clinical setting.
In the present study, we identified that different studies were controversial in their results regarding the fact that PD-L1 expression predicts a better or worse prognosis in relation to OS. However, our meta-analysis indicated that PD-L1 expression was associated with worse OS and had no significant association with DFS, which is in agreement with two other meta-analyses [1,11]. The exact mechanism between tumor and immune microenvironment remains undetermined, but new biomarkers such as CD8 and FOXP3 may contribute to the stratification of patients and a better understanding of these survival curves [48]. Another important point involves the heterogeneity of tumor PD-L1 expression and its metastatic sites, whether in axillary lymph nodes [49] or distant metastases [28,50]. Women with PD-L1-negative primary breast tumors who developed metastatic disease with PD-L1 expression seem to improve their prognosis, which favors the inclusion of this variable in new studies [28,50]. Furthermore, although the patients included in this review did not receive any immunotherapy, the effect of the treatment received on the OS and DFS curves cannot be excluded. Finally, the standardization of PD-L1 measurement by IHC and the constant training of pathologists may, in the near future, allow new associations between the expression of checkpoint inhibitors and oncological outcomes in BC [42][43][44].

Strengths and Limitations
This study presents the latest evidence of PD-L1 expression in IBC. In contrast to previous systematic reviews [1,11], only cases involving PD-L1 upregulation were considered. This study analyzed data from four major health science databases, and the review was conducted with scientific rigor. Moreover, studies that evaluated specific cases of BC (HER2-positive and triple-negative), which could confound the results, were not included in the analysis.
However, this study has some limitations. First, the search included only studies published in English from January 2018 to January 2021. The exclusion of non-English literature might have led to selection bias. Second, there was high heterogeneity between studies, which could be explained by differences in sample size, lack of standardized criteria for the immunohistochemical analysis of PD-L1, and the use of tissue microarrays. Future studies should use whole-tissue sections to evaluate TCICs according to clinicopathological features.
Recent data suggest that patients whose tumors overexpress PD-L1 have better clinical outcomes with immunotherapy [4,5,7,8]. In future research, intelligent clinical trials testing new combinations of immunotherapy and extensive evaluations of biomarkers are expected to be published, which could extend the current indications for immunotherapy in the management of BC. Furthermore, we suggest that further studies investigate whether the prognostic role of PD-L1 expression is different for BC patients with different therapies.

Conclusions
The proportion of PD-L1 expression was higher in TCICs. PD-L1 upregulation in TCs and ICs was associated with age ≥ 50 years, Ki67 index ≥ 20%, and tumors with lymph node-negative, PR-negative, or HER2-negative status. Moreover, PD-L1 upregulation was significantly associated with worse OS, but not with DFS. Funding: This work was carried out with its own financing and with partial support from Instituto Federal Goiano.