Breast cancer is the most common cancer in women, and its incidence has markedly increased in recent years. However, most patients are actually diagnosed with early-stage disease, and can thus be cured by locoregional and adjuvant systemic treatments. In recent decades, progress and intensification of adjuvant therapies in early breast cancer has significantly contributed to reducing relapse rates and improving survival.

Beyond classical prognostic factors (eg, age, tumor size, lymph node involvement, histological grade, ER, Ki67, and HER2 expression) used for treatment-related decisions, the development of multigene expression technologies in the early 2000s led to the discovery of transcriptomic heterogeneity among breast cancers, and brought about a refined tumor molecular classification with distinct clinical prognosis.1 Numerous multigene expression assays have been developed to help clinicians to refine the evaluation of the risk of relapse.2–4 Most are based on cancer cell intrinsic biology, with a dominant weight given to proliferation-associated genes.5 Among these assays, the PAM50 assay, which was initially developed to identify the different intrinsic breast cancer subtypes (luminal A/B, HER2 enriched, and basal like),3 6 is able to refine the outcome prediction, not only in ER+, but also in ER− tumors.6

Besides the intrinsic biology of breast cancer cells, recent studies on the tumor microenvironment provide new insights into the prognostic importance of immune response in breast cancer. Breast cancer was largely considered as non-immunogenic for many decades. However, the prognosis of breast cancer now appears to be associated with some features of host immune response, like tumor infiltrating lymphocyte (TIL) density, composition, and activity.7–10 Numerous studies over the past years have highlighted that the presence of high levels of TILs, but also of Th1 immune signatures, is associated with improved survival in early stage breast cancer, especially HER2-overexpressing and triple negative subtypes.11 12 Regarding the latter subtype, it was recently shown that the organization of the immune response within the tumor is also associated with prognosis.13–15 In contrast, a strong tumor infiltration in immunoregulatory cells such as regulatory T lymphocytes (Tregs), M2 macrophages, or a Th2 polarization of the immune response have been associated with a poor prognosis in early breast cancer.12 16 17

Moreover, accumulating data also suggest that the clinical efficacy of cytotoxic therapies commonly used by oncologists (and assessed by pathological complete response after neoadjuvant chemotherapy), may sometimes be positively influenced by local T dependent immune response, here again, mainly in HER2 positive and triple negative subtypes.9 10 18 In the recent context of the arrival of immunotherapy in triple negative breast cancer, a strong tumor infiltration of TILs, but also the presence of an interferon-γ (IFN-γ) inflammatory signature, or a high tumor mutational burden (TMB), have been associated with greater effectiveness of not only immunotherapy but also chemotherapy.19 These results raised the question whether certain breast tumors might benefit more from immune-based interventions and which cancer cell-intrinsic and/or microenvironmental factors define anti-tumor immune response.20

However, it must be noted that all these pioneering studies suffered from technical limitations, only evaluating the global pool of TILs (commonly detected in H&E on stained histological slides via light microscopy), or only one or two particular subtypes of immune cells, detected by specific staining in immunohistochemistry. For instance, with their role as final effectors of immune response, a favorable impact of breast tumor-infiltrating CD8+ T cells on survival has been largely reported.21 22 Nonetheless, all these histological studies offered only a narrow view of the complexity of immune response, immune cell interactions, and immune cell function and diversity in breast tumors. For example, it is still unknown whether the respective proportions of the different subpopulations of CD8+ T cells vary across intrinsic molecular subtypes of breast cancer (based on a transcriptomic definition), and how these different populations could influence early breast cancer prognosis. Moreover, in addition to CD8+ T cells, breast cancers are usually infiltrated by many other lymphoid or myeloid immune cells. This leukocyte complexity, and the respective role of each immune cell population on breast cancer prognosis, is only starting to be described.23 24 Thanks to deconvolution algorithms and public transcriptome datasets, we were able to estimate from tumor bulk mRNA the relative abundance of distinct subsets of immune and non-immune cells, and their relationships with outcome.

In order to reconcile the prognostic information given by both molecular intrinsic subtype (assessed by PAM50), and tumor global immune contexture (TIC) of early breast cancers, we performed complementary in silico analyses in more than 2800 early breast cancer transcriptomes with corresponding clinical annotations. We therefore developed a new gene expression deconvolution algorithm derived from Cibersort,25 and which accurately estimates in tumors not only the respective proportions of immune cells, but also the absolute quantity of distinct immune cell populations in the tumor sample, allowing us to study their statistical associations with relapse-free and overall survival (OS) among the different intrinsic molecular subtypes (defined by PAM50 expression assay). Moreover, this approach gave us the opportunity to study the added value of immune contexture on top of clinical parameters, and tumor molecular intrinsic subtype, for early breast cancer risk stratification.

Because of their role as effectors in the cytotoxic response against cancer cells, we first investigated whether CD8+ T cell subpopulations were different in ER-positive and ER-negative breast cancers. Figure 2 shows the estimated proportions (figure 2A), and absolute quantities (figure 2B) of CD8 fractions: effector memory CD8 T cells (EM CD8) constitute the majority of CD8 infiltrating breast tumors. Among all CD8 T cells, the estimated proportions of EM CD8 are significantly higher in ER− tumors (p value Wilcoxon test 9e-06), while the proportion of naïve CD8 is higher in ER+ tumors (p value Wilcoxon test 1e-11). There was no difference in proportions as regards central memory (CM) CD8 between ER+ and ER− tumors (p value Wilcoxon test 0.19). Similar trends were observed for the absolute quantities: compared with ER+ tumors, ER− breast cancer tissue contained more CD8 T cells, especially more EM CD8 (p value Wilcoxon test 6e-06), while ER+ tumors contained more naïve CD8+ (p value Wilcoxon test 5.7e-05). Here again, there was no difference in terms of CM CD8 (p value Wilcoxon test 0.54).

In view of these differences, we then investigated CD8 T cell subpopulations according to breast cancer molecular subtype, defined by PAM50, and found marked differences (figure 2C,D). HER2-enriched and basal-like molecular subtypes contained higher quantities of EM CD8, but also CM CD8 T cells compared with luminal tumors (Wilcoxon test basal vs luminal, p value=5e-9 and HER2 vs Luminal, p value=2e-4). Among luminal tumors, luminal B subtypes contained more EM CD8 T cells compared with luminal A tumors. Here again, among all CD8 T cells, the estimated proportions of the different CD8 subpopulations show the same results as the absolute quantities between the four molecular subtypes (figure 2D). The largest difference between pam50 subgroups is on the EM CD8 T cells whereas the differences in quantity for the CM CD8 T cells and the naïve CD8 T cells are rather small. According to their specific location in peripheral blood,32 we did not find the significant presence of intratumoral EMRA cells.

Thanks to a new gene expression deconvolution algorithm derived from Cibersort, allowing estimation of the relative and absolute quantities of various tumor-infiltrating immune cells, our study highlights the additional prognostic value of overall TIC on top of the breast cancer molecular intrinsic subtypes defined by the PAM50 classifier.

Our results show that the different CD8 T cell subpopulations, but also more generally, the immune contexture of breast tumors, as assessed by hierarchical clustering based on both myeloid and lymphoid cells, differ between breast cancer molecular intrinsic subtypes. They also have variable and complex associations with different level of clinical impact in terms of patient outcome. Finally, our findings make it possible to refine the survival stratification of patients with early breast cancer by incorporating TIC with PAM50 and clinical tumor burden (T and N stage) in a prognostic model assessed in training and validation cohorts.

The four breast cancer intrinsic subtypes were first described by Perou et al,1 and were initially defined by a large set of genes, subsequently reduced to 50 classifier genes in the PAM50 assay.6 36 Numerous studies have shown that PAM50 provides better estimates of early breast cancer patient outcome than classical clinicopathological factors.3 37 38 Moreover, PAM50 intrinsic subtypes, especially PAM50 proliferation genes, have also been shown to have a predictive value concerning the benefit of chemotherapy.6 38–41 Indeed, among the 50 genes involved in PAM50, the majority are related to proliferation, estrogen, or other growth factor pathways, but are not linked to immune response.6

Breast cancer was previously thought to be a non-immunogenic cancer, but numerous recent studies have suggested that in some cases, the presence of TILs can be a strong indicator of immune response against cancer cells, and thus a favorable prognostic factor.7 9 10 Indeed, in TNBC and HER2-overexpressing breast tumors, retrospective analyses of large adjuvant/neoadjuvant studies have shown that high level of TILs and CD8 T cells, or a Th1 cytotoxic signature, are independently associated with better survival and/or with better response to chemotherapy.7 9–11 18 42 In accordance with these pioneering histological studies, we found that TNBC and HER2-overexpressing cancers are breast tumors with the highest immune cell infiltration, and molecular subtypes in which most, if not all, immune cells are associated, or tend to be associated, with better survival. This was previously demonstrated in transcriptomic analyses from TNBC reported by Denkert et al,43 in which all immune genes were positively correlated among themselves (and with favorable treatment response), even when genes were markers of immunosuppressive populations.

More recent studies, based, like ours, on deconvolution analysis of gene expression data coming from early breast cancer, have highlighted the complexity of tumor-infiltrating immune cell types, and their respective proportions and prognostic roles.23 24 Accordingly, our unsupervised clustering analysis using immune cell quantities revealed four subgroups of breast tumors, largely independent of the PAM50 intrinsic subtype, and with distinct risks of relapse and death. These results indicate that the patterns and prognostic impact of immune infiltrates are not strictly linked to the biological tumor behavior defined by the molecular intrinsic subtypes. Accordingly, we demonstrate here the complementarity of both tumor intrinsic biology and tumor complex immune response, to better assess the prognosis of early breast cancer.

From an immunological point of view, our study demonstrates the prognostic relevance of 15 different subsets of immune cells among myeloid and lymphoid populations. More specifically, we describe here the relative infiltration of different CD8+ T cells. Among TILs, CD8+ T cells are believed to be one of the most important components of cell-mediated immunity, able to directly kill tumor cells presenting tumor-associated antigenic peptides. Accordingly, high tumor-infiltrating CD8+ cell density has been independently associated with improved survival in various types of cancer, including breast cancer.42 In breast cancer, this favorable effect appeared to be restricted to HER2-overexpressing and TNBC, with no significant impact in ER+/HER2− subtype.44

However, many different subpopulations of CD8+ T cells can be found in the tumor immune microenvironment, including naïve, memory, effector, and exhausted lymphocytes. After antigen encounter, naïve CD8+ T cells are activated and differentiate into memory T -cell populations, including effector CD8+ T cells.45 This differentiation is accompanied by the cell’s ability to proliferate and home into lymphoid organs, and to acquire a proinflammatory and cytotoxic phenotype.45 46 Among memory T cell populations, EM T cells are able to rapidly produce IFN-γ after antigenic recall stimulation, while CM T cells express lymph node homing receptors and produce less IFN-γ on stimulation.47

Our analyses revealed that EM CD8+ T cells constitute the predominant population of breast cancer CD8+ infiltrating T cells. These cells were more abundant in basal-like and HER2 molecular subtypes. Such results have been obtained elsewhere by flow cytometry analysis in a small cohort of localized breast cancer.48 Interestingly, in this translational study, and mostly based on luminal breast tumors, these CD8+ TILs expressed checkpoint molecules like programmed cell death protein 1 (PD-1) or TIGIT and are enriched in an exhausted phenotype. However, the authors demonstrated that these CD8+ infiltrating T cells retained the capacity to produce cytokines and to be cytotoxic.48 This could explain the favorable, and preponderant prognostic role of EM CD8+ T cells observed in our study. Indeed, compared with CM or naïve CD8 T cells, our results highlighted the prominent prognostic importance of effector memory cells, among the CD8 T cell population. The presence of CD45RO+ tumor infiltrating memory T cells has been associated with an absence of signs of early metastatic invasion, less advanced pathological stage, and increased survival in pioneering works conducted in localized colorectal cancer samples.49 More recently, a specific favorable role of EM CD8 T cells has been deciphered in ovarian cancer,50 and melanoma.51 In ipilimumab-treated metastatic melanoma patients, peripheral CD8 effector-memory type 1 T-cells have been correlated with more favorable outcome,52 highlighting their potential role as a predictive biomarker for immunotherapy efficacy.

Interestingly, in our breast cancer cohorts, this favorable prognostic effect was observed in basal like and HER2 tumors, but also significantly in luminal B subtypes. This is another means of underlining the complementarity of both immunological and tumor molecular intrinsic information to decipher the complexity of TIL impact in ER+ luminal breast tumor prognosis. Indeed, our results demonstrate a clear and different effect of CD8, and more generally, of the different subsets of immune cells included in TIC, among luminal A and luminal B tumors, with a clear favorable prognostic impact of lymphoid populations (but not myeloid cells, which tend to be associated with worse outcome) for luminal B tumors, whereas no significant effect can be seen for luminal A tumors. These results may partially explain the controversial role of TILs and immune response in luminal breast cancer. In this respect, it should be noted that our final prognostic model incorporating tumor stage, PAM50 and TIC prognostic information was more efficient in ER+ tumors than PAM50 alone or combined with tumor stage, making it possible to discriminate three different populations in terms of outcome, notably a group of patients with nearly 100% OS at 10 years and in contrast, a group of patients with very poor survival considering ER+ localized breast cancer. These results highlight, in our view, the clinical importance of immune response in early breast cancer, even in ER+subtypes.

It should be noted that usually, such correlative studies do not constitute formal proof of a causal relationship, leaving open questions as to whether all immune subsets of TILs cooperate to influence the prognosis of patients, or whether these associations with survival are solely the result of tumor cell intrinsic biological conditions. In this regard, our results show for the first time that different clusters of overall TIC in breast cancer appear to be independent of the intrinsic molecular subtypes usually defined by PAM50. This argues for a specific role of immune response in the prognosis of patients with breast cancer, and for the complementarity of both biological information to refine patient prognosis. This refinement in the prediction of clinical outcome makes it possible to identify, within each subtype of breast cancer, patients with good, intermediate or poor prognosis, thereby opening the door for tailored adjuvant treatment, according to the risk of relapse and death. These results are of particular interest in the recent context of new immunotherapies, since the prognostic and predictive value of antitumor immune response in breast cancer indicates that immunotherapy could constitute a promising option for the treatment of some breast cancers. However, in breast cancer, despite some clinical response observed with anti-PD-1 or anti-programmed death-ligand 1 (PD-L1) monotherapy, the rate of responders remains low compared with other cancer types (such as melanoma, kidney, or non-small cell lung cancer), even in the triple negative subtype, despite higher rates of TILs, PD-L1 expression, and higher TMB.53 A better understanding of tumor-infiltrating immune cell composition among the different breast cancer molecular subtypes could benefit the future design of specific immunotherapeutic clinical trials for patients with breast cancer in each molecular type of disease.

Our study has other limitations: thus, the immunological parameters that we integrated into our models do not take into account the degree of activation or inactivation of the different immune populations (evaluated in part by the degree of expression of the checkpoints of immune response), nor the spatial organization of these immune cells, to which we have not had access. As these two pieces of information are also associated with the prognosis of early breast cancer,12 13 15 54 it could be interesting to study to what extent these parameters could be complementary with our approach to further refine the individual prognosis.

In addition, whether our tool could also help to select patients for standard adjuvant systemic treatment (eg, endocrine treatment or chemotherapy) remains unclear. Additional studies in cohorts of patients treated with neoadjuvant strategies are ongoing, and would help us to understand whether overall TIC, and our prognostic model also enable identification of responding tumors or not (as a predictive biomarker).

In conclusion, our study provides new insights into the complexity of breast cancer immune response, and its clinical relevance, when combined with tumor intrinsic molecular subtypes to refine prognosis of patients with breast cancer in each subtype. We bring together in the context of early breast cancer, the prognostic information given by both molecular intrinsic subtype (assessed by PAM50), and by the overall TIC assessed by a new gene expression deconvolution algorithm. However, the complexity of immune infiltrate interplay, and how they influence long term outcome, remains largely unclear in breast cancer, as in the majority of other solid cancers. In the future, novel biological techniques, like single-cell molecular analyses will certainly help to reveal the profound complexity of cellular subsets, and their molecular and functional properties. Together with an improved understanding of tumor cell intrinsic biology, these techniques of high analytic resolution could be complementary to multiplexed histological analyses55 and high throughput bioinformatic tools in deciphering the complex landscape of immune response against cancer cells, and providing better rationales for the design of future clinical trials in immunotherapy and oncology.