PD-1 protein and gene expression as prognostic factors in early breast cancer.

BACKGROUND: There is a paucity of data on the prognostic value of programmed cell death protein 1 (PD-1) protein and gene expression in early breast cancer (BC) and the present study's aim was to comprehensively investigate it.
METHODS: The study consisted of three parts: a correlative analysis of PD-1 protein and gene expression from an original patient cohort of 564 patients with early BC; a systematic review and trial-level meta-analysis on the association between PD-1 protein expression and disease-free survival/overall survival (OS) in early BC; and a pooled gene expression analysis from publicly available transcriptomic datasets regarding PDCD1 expression.
RESULTS: In the study cohort, PD-1 protein, but not gene expression, was associated with improved OS (HRadj=0.73, 95% CI 0.55 to 0.97, p=0.027 and HRadj=0.88, 95% CI 0.68 to 1.13, p=0.312, respectively). In the trial-level meta-analysis, PD-1 protein expression was not found to be statistically significantly associated with outcomes in the overall population. Finally, in the pooled gene expression analysis, higher PDCD1 expression was associated with better OS in multivariable analysis in the entire population (HRadj=0.89, 95% CI 0.80 to 0.99, p=0.025) and in basal-like tumours.
CONCLUSIONS: PD-1 protein and gene expression seem to be promising prognostic factors in early BC. Standardisation of detection and assessment methods is of utmost importance. © Author (s) (or their employer(s)) 2020. Re-use permitted under CC BY-NC. No commercial re-use. Published by BMJ on behalf of the European Society for Medical Oncology.

The prognostic role of programmed cell death protein 1 (PD-1) protein and gene expression in breast cancer is not well studied. Even fewer studies report on their association which remains unclear owing to analytical difficulties and the complex regulation of the PD-1/programmed cell death ligand 1 axis.

PD-1 protein expression predicted better survival in an original patient cohort, though this finding was not confirmed in a trial-level meta-analysis of 15 studies. In addition, PD-1 gene expression predicted better survival in a pooled gene expression analysis. The prognostic effect was mainly driven by triple negative/basal-like breast cancer. PD-1 protein and gene expression were weakly, but significantly, correlated.

Emerging data show that PD-1 gene expression might be a predictor of benefit from immune checkpoint inhibition. This study confirms its prognostic role and supports its future inclusion to management algorithms as a marker that selects a patient population, mainly with triple negative disease, with inherently better prognosis that may additionally benefit from immunotherapy.

The prognostic role of programmed cell death protein 1 (PD-1) protein and gene expression in breast cancer is not well studied. Even fewer studies report on their association which remains unclear owing to analytical difficulties and the complex regulation of the PD-1/programmed cell death ligand 1 axis.

PD-1 protein expression predicted better survival in an original patient cohort, though this finding was not confirmed in a trial-level meta-analysis of 15 studies. In addition, PD-1 gene expression predicted better survival in a pooled gene expression analysis. The prognostic effect was mainly driven by triple negative/basal-like breast cancer. PD-1 protein and gene expression were weakly, but significantly, correlated.

Emerging data show that PD-1 gene expression might be a predictor of benefit from immune checkpoint inhibition. This study confirms its prognostic role and supports its future inclusion to management algorithms as a marker that selects a patient population, mainly with triple negative disease, with inherently better prognosis that may additionally benefit from immunotherapy.

Introduction

While the era of immunotherapy of breast cancer (BC) through checkpoint inhibition1 has begun with the IMPassion130 trial,2 the quest to identify clinically useful biomarkers is far from over. The most widely used is protein expression of programmed cell death ligand 1 (PD-L1) as assessed by immunohistochemistry (IHC). Analytical difficulties and the demonstration that other biomarkers may more accurately predict benefit from immunotherapy are causes for concern.3 4 In addition, the improvement in pathological complete remission rates regardless of PD-L1 status in the KEYNOTE-522 trial5 further supports these concerns. At the same time, while the prognostic value of PD-L1 protein expression in BC remains unclear, PD-L1 gene expression is a marker clearly associated with good prognosis, whose potential clinical validity we have previously demonstrated.6 7 These observations pose the question whether results from IHC evaluation of PD-L1 are assay driven, rather than representative of the underlying biology.

Due to the complexity of programmed cell death protein 1 (PD-1)/PD-L1 axis regulation,8 other predictive biomarkers have been explored. Of those, high PD-1 mRNA expression, thought to represent a stable pre-existing adaptive immune response,4 is particularly promising and may outperform other markers, including PD-L1 IHC, as a predictor of benefit from PD-1/PD-L1 inhibition.4 The analytical validity and reproducibility of PD-1 mRNA evaluation have been previously demonstrated,9 underscoring the potential of this marker.

To our knowledge, few studies have reported on the prognostic value of PD-1 protein expression in BC, while available data on PD-1 gene expression are even more sparse. Herein, we aimed to investigate the prognostic value of PD-1 protein and gene expression in a cohort of patients treated in Stockholm, Sweden. In addition, in order to interpret our findings in the context of current evidence, we performed a trial-level meta-analysis of available data on PD-1 IHC in BC and a separate pooled analysis of transcriptomic data from publicly available datasets.

Methods

Three parts comprised the present study: a correlative analysis of PD-1 protein and gene expression from an original patient cohort; a trial-level meta-analysis on PD-1 protein expression in early BC; and a pooled gene expression analysis from publicly available datasets regarding PDCD1 expression. The corresponding methods of the three parts are presented hereunder in succession.

Description of the study cohort

The study cohort consisted of patients treated for primary BC in Stockholm, Sweden during 1997–2005 and has been previously described in detail7 10 (online supplemental figure S1). The median follow-up for overall survival (OS), defined as time from date of diagnosis to death by any cause, was 15.7 years and the median follow-up for distant recurrence-free interval (DRFI), defined as time from date of diagnosis to distant metastasis or death from BC,11 was 12.6 years. The used methods and strategies in the present study are covered by the ethics application including amendments, approved by the ethics committee at Karolinska Institutet, Stockholm, Sweden.

Tissue microarray construction, staining and scoring for PD-1 protein expression

Tissue microarray (TMA) construction was performed as previously described.7 Immunostaining with the primary monoclonal antibody targeting PD-1 (clone NAT105, Roche, Basel, Switzerland) was performed using the Ventana (Roche, Basel, Switzerland) autostainer according to the manufacturer’s protocol. Reactive lymphoid node of the tonsil was used as a positive control. PD-1 was evaluated by two investigators (IZ, GR), while TMA cores with poor staining quality, folded, limited or missing tissue were excluded from the analysis. The total number of PD-1 positive (PD-1+) immune cells—defined as the presence of any membranous staining—was manually counted in each TMA core and then calculated between duplicate cores; thus, the average number of PD-1+ immune cells per sample was derived. Any staining of PD-1 +immune cells was used as cut-off for positivity.

Gene expression profiling

RNA was extracted from the obtained biopsies or surgical specimens and profiled on Rosetta/Merck Human RSTA Custom Affymetrix V.2.0 microarray as previously reported and are available at the Gene Expression Omnibus database under accession number GSE48091.7 Assignment of the intrinsic subtype in each tumour according to the PAM50 classification12–14 has been described previously in detail.10

Statistical methods for the PD-1 protein and gene expression analyses in the study cohort

In this exploratory study, association between PD-1 protein and mRNA expression levels and DRFI and OS was analysed. Kaplan-Meier curves were used to estimate the survival outcomes and dichotomised PD-1- protein levels (positive, negative) were compared with the non-parametric log-rank statistic. Univariate and multivariable Cox regression models with dichotomised PD-1 protein level as categorical variable were fitted and unadjusted and adjusted HRs and CIs were calculated. In the multivariable Cox regression models, adjustment was made for age (<50, 50+years), tumour size (≤2, 2–5,>5 cm), lymph node involvement (yes, no), histological grade (I, II, III), proliferation status (Ki67 ≤20%,>20%) and IHC-based subtype (ER+/HER2−, HER2+, ER−/HER2−). Missing values of adjustment variables were treated with separate categories in the models. Association between PD-1 mRNA expression levels and outcome was analysed with corresponding Cox regression models with PD-1 mRNA expression value as continuous variable after robust linear scaling (such that expression quantiles 2.5% and 97.5% were set to −one and +1, respectively). All statistical analysis was done in R (V.3.6.2).

Search strategy, study selection, data extraction and quality assessment in the meta-analysis of PD-1 protein expression

Comprehensive systematic electronic search was conducted by a librarian at the Karolinska Institutet University Library in November 2018 and updated in May 2019 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 search strategy has been described in detail6 and provided in the online supplemental material.

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 assessment, tissue used for analysis, threshold for PD-1 expression, antibody used; positivity rate of PD-1 in immune cells; characteristics of study cohort; follow-up time; outcome (time-to-event variables) within all patients and whenever possible within different BC subtypes including both univariate and multivariate results.

Only studies investigating the prognostic role (measured as time-to-event outcome) of PD-1 expression in patients with early-stage BC were included in our meta-analysis. Study selection was performed independently by two investigators (AM, IZ) and consensus was reached in all eligible studies. Data were independently extracted in a predefined form. A third investigator (AV) compared the databases and resolved any discrepancies. The concordance rate between the two investigators was 97.8%.

Two investigators (AM, IZ) independently assessed each eligible study for methodological quality using the 20-item REMARK (Reporting recommendations for tumour marker prognostic studies) checklist, as previously described.15 Discrepancies were resolved by a third investigator (AV). The concordance rate between the two investigators considering quality assessment was 70.0%.

Statistical analysis for study-level meta-analysis of PD-1 protein expression

Each case was defined as positive when PD-1 expression exceeded the positivity threshold used in the specific study. For the time-to-event outcomes, 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 et al.16 Whenever possible, we performed separate analyses for univariate and multivariate HRs in each outcome of interest. Subgroup analyses were performed if there were at least three studies in each subgroup.

We assessed the presence of statistical heterogeneity among the studies using the Q statistics and the magnitude of heterogeneity using the I2 statistic. We considered a p value<0.10 or an I2 value of greater than 50% as indicative of substantial heterogeneity. If substantial heterogeneity was detected, we used the random-effect model; otherwise, the fixed-effect model was used.

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 V.5.3 (The Cochrane Collaboration, 2014).

Gene expression data and statistical methods for pooled gene expression analysis

Thirty-nine datasets of gene expression profiles of nearly 9500 primary BC were used in the pooled transcriptomic analysis, as previously described.6 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 tumour was assigned to a molecular intrinsic subtype according to the PAM50 classifier. Tumours of the normal-like subtype were not included in the subgroup analysis. ESR1 and ERBB2 status were classified based on the bimodal distribution of the expression values of these two genes, since not every dataset had information on ER and HER2 status.

Association between PD-1 mRNA expression levels and disease-free survival (DFS, DMFS [distant metastasis-free survival], RFS [relapse-free survival] or progression-free interval) and OS were analysed. Univariate and multivariable Cox regression models with scaled expression value as continuous variable and stratification by cohort were fitted, and unadjusted and adjusted HR and CI were calculated. In the multivariate Cox regression model, adjustment was made for age, tumour size, lymph node involvement, histological grade and ESR1 and ERBB2 expression status. No adjustment for ESR1 and ERBB2 status was made in models by molecular subtype. Each METABRIC (Molecular Taxonomy of Breast Cancer International Consortium) study site (n=5) was treated as separate cohort. All data analysis was done in R/Bioconductor (V.3.6.2).

Results

Patient characteristics in the study cohort

The patients’ clinicopathological characteristics in the entire study cohort and according to PD-1 protein (positive vs negative; online supplemental figure S2) are summarised in table 1. In total, 220 (46.7%) had PD-1 protein positive expression and 251 (53.3%) negative. PD-1 protein expression was associated with biologically aggressive characteristics, such as oestrogen receptor negativity (p<0.001) and high proliferation according to Ki67 (p<0.001).

Table 1

Patient characteristics for all patients in study cohort and according to programmed cell death protein 1 (PD-1) protein expression (positive vs negative)

Clinicopathological characteristic	All	Patients with PD-1 IHC dataN (%)	PD-1 positiveN (%)*	PD-1 negativeN (%)*	P value†
Number of patients	564	471 (100.0)	220 (46.7)	251 (53.3)
Age					0.795
Median in years (IQR)	54.5 (23–76)	54 (23–76)	55 (23–75)	53 (23–76)
Tumour size					0.799
Median in cm (IQR)	2.20 (0.20–12.30)	2.20 (0.20–12.30)	2.20 (0.70–9.50)	2.20 (0.20–12.30)
Lymph node status					0.760
Positive	313 (55.5)	271 (57.5)	125 (56.8)	146 (58.2)
Negative	234 (41.5)	186 (39.5)	87 (39.5)	99 (39.4)
Unknown	17 (3.0)	14 (3.0)	8 (3.6)	6 (2.4)
ER status					<0.001
Positive	399 (70.7)	329 (69.9)	133 (60.5)	196 (78.1)
Negative	152 (27.0)	129 (27.4)	77 (35.0)	52 (20.7)
Unknown	13 (2.3)	13 (2.8)	10 (4.5)	3 (1.2)
PR status					0.005
Positive	270 (47.9)	223 (47.3)	87 (39.5)	136 (54.2)
Negative	152 (27.0)	131 (27.8)	73 (33.2)	58 (23.1)
Unknown	142 (25.2)	117 (24.8)	60 (27.3)	57 (22.7)
HER2 status
Positive	97 (17.2)	88 (18.7)	39 (17.7)	49 (19.5)	0.888
Negative	385 (68.3)	338 (71.8)	160 (72.7)	178 (70.9)
Unknown	82 (14.5)	45 (9.6)	21 (9.5)	24 (9.6)
Tumour grade					0.088
Grade I	46 (8.2)	34 (7.2)	17 (7.7)	17 (6.8)
Grade II	244 (43.3)	210 (44.6)	85 (38.6)	125 (49.8)
Grade III	245 (43.4)	221 (46.9)	114 (51.8)	107 (42.6)
Unknown	29 (5.1)	6 (1.3)	4 (1.8)	2 (0.8)
Ki67					<0.001
≤20%	230 (40.8)	212 (45.0)	120 (54.5)	92 (36.7)
>20%	289 (51.2)	253 (53.7)	96 (43.6)	157 (62.5)
Unknown	45 (8.0)	6 (1.3)	4 (1.8)	2 (0.8)
IHC subtype					0.008
ER+/HER2−	283 (50.2)	248 (52.7)	103 (46.8)	145 (57.8)
HER2+	97 (17.2)	88 (18.7)	39 (17.7)	49 (19.5)
ER−/HER2−	93 (16.5)	81 (17.2)	51 (23.2)	30 (12.0)
Unknown	91 (16.1)	54 (11.5)	27 (12.3)	27 (10.8)
PAM50-based subtype					0.001
Luminal A	199 (35.3)	163 (34.6)	67 (30.5)	96 (38.2)
Luminal B	112 (19.9)	98 (20.8)	37 (16.8)	61 (24.3)
HER2-enriched	78 (13.8)	69 (14.6)	30 (13.6)	39 (15.5)
Basal-like	135 (23.9)	113 (24.0)	71 (32.3)	42 (16.7)
Normal-like	40 (7.1)	28 (5.9)	15 (6.8)	13 (5.2)

*Percentages calculated based on the number of patients with available data.

†P values from statistical tests where unknown categories are excluded. Wilcoxon rank sum test is used for the continuous variables and Fisher’s exact test is used for categorical variables.

ER, oestrogen receptor; HER2, human epidermal growth factor receptor 2; IHC, immunohistochemistry; PR, progesterone receptor.

Prognostic value of PD-1 protein expression and its correlation with PD-1 mRNA expression in the study cohort

In the study cohort, high PD-1 protein but not mRNA expression was generally found to be prognostic for improved patient outcomes. In multivariable analysis, PD-1 protein expression was associated with better DRFI (adjusted HR (HRadj)=0.66, 95% CI 0.48 to 0.91, p=0.010) and OS (HRadj=0.73, 95% CI 0.55 to 0.97, p=0.027; figure 1 and online supplemental table S1). In subgroup analysis, this prognostic effect was mainly driven by ER-positive/HER2-negative disease (for DRFI, HRadj=0.65, 95% CI 0.41 to 0.1.01, p=0.050; for OS, HRadj=0.70, 95% CI 0.47 to 1.04, p=0.076), although there was no interaction between PD-1 expression and IHC-based subtype in the survival analyses (figure 1; pinteraction 0.326 for DRFI and pinteraction 0.443 for OS).

Figure 1

Kaplan-Meier curves and univariate HRs for distant metastasis-free survival (A) and overall survival (B) and forest plots for distant recurrence-free interval (C) and overall survival (D) in the overall populations and per immunohistochemistry-based subtype (multivariable HR), according to programmed cell death protein 1 (PD-1) protein expression (positive vs negative). HER2, human epidermal growth factor receptor 2; IHC, immunohistochemistry.

In contrast, high PD-1 mRNA expression was not prognostic in multivariable analysis for neither DRFI (HRadj=0.88, 95% CI 0.66 to 1.18, p=0.399), nor OS (HRadj=0.88, 95% CI 0.68 to 1.13, p=0.312) (online supplemental table S1). PDCD1 transcript levels were statistically significantly but weakly correlated with PD-1 protein expression (Spearman’s rho=0.123, p=0.007).

Characteristics and quality of eligible studies and between-study heterogeneity in the meta-analysis of immunohistochemical PD-1 expression

As shown in figure 2, the initial search identified 4736 entries, or 3298 entries following deduplication. Through exclusion by reading the title and/or abstract, 39 possibly eligible studies were retrieved as full text; 15 studies fulfilled the inclusion criteria and were included in the meta-analysis.17–30 The characteristics of eligible studies are presented in online supplemental table S2. The results from the study cohort described in the Methods section and presented in the Results section were included in the meta-analysis as one of the 15 eligible studies.

Figure 2

Flow chart of search and study selection in the meta-analysis of studies evaluating protein programmed cell death protein 1 (PD-1) expression (A); and data availability in the pooled gene expression analysis (B). DFS, disease-free survival; OS, overall survival.

Fourteen studies were retrospective and one prospective. The median number of study quality score was 23.5 (range: 19–33) out of a maximum score of 40. Substantial between-study heterogeneity was noted among eligible studies regarding the threshold for PD-1 positivity rate used, the follow-up period, and the study population and BC subtypes.

Prognostic significance of pooled immunohistochemical PD-1 expression

PD-1 positive disease was not significantly associated with OS in the overall population (figure 3). The pooled (7 studies; 1975 patients) univariate HR was 1.65 (95% CI 0.94 to 2.90, p=0.08) and the pooled (four studies; 1395 patients) multivariate HR was 1.25 (95% CI 0.71 to 2.20, p=0.45). Significant heterogeneity was noted in both analyses (I=88, pheterogeneity <0.001 and I=78%, pheterogeneity=0.004, respectively). Similar results were observed for DFS (5 studies, n=633; pooled univariate HR=1.71, 95% CI 0.95 to 3.08, p=0.07; I=57%, pheterogeneity=0.05).

Figure 3

Forest plot for overall survival according to programmed cell death protein 1 (PD-1) protein expression in total population: pooled univariate HR (A) and pooled multivariable HR (B).

The association of PD-1 positivity and improved DFS in TNBC was statistically significant (7 studies, n=1245; pooled univariate HR=0.52, 95% CI 0.38 to 0.73, p<0.001; I=52%, pheterogeneity=0.05 and 3 studies, n=417; pooled multivariate HR=0.57, 95% CI 0.29 to 0.90, p=0.02, I=12%, pheterogeneity=0.32) but not with OS in pooled univariate (9 studies, n=1454; HR=0.76, 95% CI 0.44 to 1.31, p=0.32, I=84%, pheterogeneity <0.001) and pooled multivariate analyses (3 studies, n=437; HR=0.74, 95% CI 0.47 to 1.17, p=0.20, I=8%, pheterogeneity=0.34; online supplemental figure S3). Due to the limited number of studies, further analyses were not possible.

Pooled analysis of PD-1 gene expression as a prognostic factor

As previously described,6 39 datasets containing 9493 patients were included in the pooled gene expression analysis. The prognostic value of PD-1 gene expression in univariate and multivariable analysis is shown in figure 4 and in online supplemental tables S3 and S4). Higher PDCD1 expression was associated with better OS in multivariable analysis in the entire population (HRadj=0.89, 95% CI 0.80 to 0.99, p=0.025) and in basal-like (HRadj=0.77, 95% CI 0.63 to 0.95, p=0.014) as well as HER2-enriched tumours (HRadj=0.78, 95% CI 0.61 to 0.98, p=0.032), although no significant interaction between PD-1 expression and molecular subtype was detected (pinteraction=0.185). In contrast, PDCD1 expression was not predictive for DFS as shown in online supplemental table S3.

Figure 4

Forest plot for overall survival according to programmed cell death protein 1 (PD-1) gene expression, in total population and per molecular intrinsic subtype. HER2, human epidermal growth factor receptor 2.

Discussion

By using multiple approaches (IHC, gene-expression profiling data, trial-level meta-analysis and pooled transcriptomic analysis), we demonstrate in this study that PD-1 expression at the protein and mRNA levels seem to be associated with better prognosis in early BC, especially—but not exclusively—in the triple-negative and basal-like subtypes. These findings are in line with our previously published study on the correlation of PD-L1 gene and immune cell protein expression with improved outcomes,6 further highlighting the prognostic significance of the PD-L1/PD-1 axis and the necessity for biomarker development.3 31

The favourable survival outcomes with higher PD-1 expression might seem paradoxical given that it is considered to be a coinhibitory molecule. However, it is not certain if PD-1 represents a T-cell exhaustion marker, thus leading to impaired antitumour immune response, or a marker indicating T-cell activation and receptor signalling.32 Indeed, an association of PD-1 expression with genes reflecting T-cell regulation and activation status has been described.33 Of note, additional immune checkpoint markers (ie, CD39, LAG-3, TIM-3) can also define T-cell exhaustion34 and are linked with better outcomes.35 Moreover, the improved patient outcomes among high expressors may also reflect the potential predictive value of PD-1 in patients receiving chemotherapy, and thus higher benefit from adjuvant therapy, in accordance with previous data showing that immune-related gene expression is a driver of chemosensitivity in BC.36 37

Even though improved OS in PD-1high vs PD-1low triple-negative/basal-like patients was observed in the pooled analyses with no benefit in DFS owing likely to the exploratory nature of this study, a discrepancy in the prognostic information between protein and mRNA levels was noted in our study cohort. Regarding protein expression, the lack of survival benefit in ER-positive/HER2-negative BC in the meta-analysis could be due to the short follow-up of these studies compared with our cohort masking a possible long-term effect of immune activation on tumour dormancy.38 The observed weak correlation between PD-1 protein and transcript levels raises concerns on how informative PD-1 IHC evaluation really is. Besides analytical difficulties, this discrepancy might be related to the complex regulation of PD-1 expression, especially at the post-transcriptional/translational level.8 39 Data on PDCD1 expression are scarce, with few published studies demonstrating an association with improved survival outcomes.29 33 These observations further underscore the need for validation of the genomic platforms for the detection of PD-1,9 while its clinical utility regarding both its prognostic and predictive4 role needs to be tested in a prospective manner.

This study suffers from certain limitations that need to be addressed. First, the meta-analysis was performed at a trial level, with relatively few retrospective studies identified and included, which did not allow for further subgroup analyses of potential interest and led to significant heterogeneity in some of the analyses. Second, variations in IHC scoring, antibodies used and cut-offs, mRNA detection methodology and patient selection, as well as variations in defining ER and HER2 positivity also contributed to the high heterogeneity among the studies. In addition, the smaller size, different patient population of our original study and non-randomised allocation to adjuvant chemotherapy, as well as PD-1 IHC evaluation with different cut-off and on TMA rather than in whole sections, which possibly underestimates the level of spatial heterogeneity,6 might explain the observed discrepant results as compared with the pooled analyses. The risk for bias and the limitations regarding analyses of publicly available transcriptomic datasets have been also described elsewhere.40 Finally, our study cannot provide information on the effect of systemic therapy in either the study cohort or the meta-analysis studies or the pooled gene expression analysis, since allocation to adjuvant therapy was not randomised.

Given the emerging impact of PD-L1/PD-1 signalling axis, its complex regulation37 39 and the early translational research data on immune checkpoint blockade in BC,4 it is of utmost importance to rethink and better understand its impact in the context of tumor-immune microenvironment. Multiplexed imaging and multiomics approaches31 can further delineate the expression patterns and specific immune cell composition as well as their prognostic and/or predictive value. Considering the possible predictive value of PD-1 mRNA for immunotherapy across tumour types and the inconsistent results of PD-L1 as a predictive marker in BC in both the neoadjuvant5 41 and metastatic settings,2 42 PD-1 mRNA may very well represent a potential candidate for integration into future treatment algorithms of both early and advanced BC as a fourth biomarker besides ER, PR and HER2.

In conclusion, this multilevel study reveals the promising prognostic capacity of PD-1 expression in patients with early BC, especially in the triple-negative and basal-like subtypes. However, the low concordance between protein and gene expression and analytical inconsistencies in reported methodologies of PD-1 IHC assessment underscore the need for prospective validation of our findings.