Multiplexed immunofluorescence identifies high stromal CD68+PD-L1+ macrophages as a predictor of improved survival in triple negative breast cancer

Triple negative breast cancer (TNBC) comprises 10–15% of all breast cancers and has a poor prognosis with a high risk of recurrence within 5 years. PD-L1 is an important biomarker for patient selection for immunotherapy but its cellular expression and co-localization within the tumour immune microenvironment and associated prognostic value is not well defined. We aimed to characterise the phenotypes of immune cells expressing PD-L1 and determine their association with overall survival (OS) and breast cancer-specific survival (BCSS). Using tissue microarrays from a retrospective cohort of TNBC patients from St George Hospital, Sydney (n = 244), multiplexed immunofluorescence (mIF) was used to assess staining for CD3, CD8, CD20, CD68, PD-1, PD-L1, FOXP3 and pan-cytokeratin on the Vectra Polaris™ platform and analysed using QuPath. Cox multivariate analyses showed high CD68+PD-L1+ stromal cell counts were associated with improved prognosis for OS (HR 0.56, 95% CI 0.33–0.95, p = 0.030) and BCSS (HR 0.47, 95% CI 0.25–0.88, p = 0.018) in the whole cohort and in patients receiving chemotherapy, improving incrementally upon the predictive value of PD-L1+ alone for BCSS. These data suggest that CD68+PD-L1+ status can provide clinically useful prognostic information to identify sub-groups of patients with good or poor prognosis and guide treatment decisions in TNBC.


Results
TNBC patient cohort. The clinicopathological features of the patient cohort (n = 244) are as previously published 28 and are summarised as follows (Supplementary Table S1 online): the average tumour size was 25.9 mm (range 7-120 mm), with patient age ranging from 25.9 to 96 years old. Median follow-up length was 4.3 years (range 0.02-16.3 years) for overall survival (OS, death from any cause) and breast cancer specific survival (BCSS, death directly attributable to breast cancer). 232 cases (95.1%) were grade 3, 85 cases (34.8%) had node-positive disease, and 111 (45.5%) had sTILs > 30%. 174 (71.3%) patients received adjuvant chemotherapy (regimens included cyclophosphamide, methotrexate 5-fluorouracil; anthracycline, cyclophosphamide; anthracycline, cyclophosphamide, paclitaxel; 5-fluorouracil, epirubicin, cyclophosphamide). Eighty six of 174 (49%) patients received a regimen which contained a taxane. Histologically, 221 (90.6%) were invasive ductal carcinoma of no special type with 17 (7%) metaplastic carcinomas and 6 (2.5%) other (apocrine, micropapillary, lobular). There were 71 deaths, 48 of which were breast cancer related. TNBC status was defined using current guidelines outlined by the College of American Pathologists (≤ 1% staining for ER and PR and HER2 negativity by IHC or silver in-situ hybridisation). All cases were scored for stromal TILs on the corresponding whole tumour section by an experienced breast Pathologist using standardised criteria 29 . Identification of specific stromal immune cell phenotypes. Using mIF, we were able to detect immunophenotypes defined by expression of up to four different markers predominantly within the stroma (Fig. 1). From the observed immunophenotypes, we investigated only those with the most clinical relevance, excluding inappropriate marker combinations (e.g., CD68 + FOXP3 + ) and those with inadequate data points (less than one cell detected). The final immunophenotypes investigated and their median cell count densities (used as the cut-point value for subsequent analyses) are listed in Table 1, along with cell count distributions in Fig. 2 www.nature.com/scientificreports/ A bar chart summarising PD-L1 status according to immune marker phenotype in presented in Supplementary  Fig. S1. Intra-epithelial cell counts for many samples were less than 1 and therefore not further assessed. Values from control normal breast tissue samples from a single patient are also included for comparison (Supplementary Table 4).

Discussion
Despite its overall poor survival, a sub-group of TNBC patients have a good prognosis and respond well to standard of care chemotherapy, often correlating with high stromal TILs levels. Current guidelines for TILs assessment however do not use immunophenotypic data to assess prognostic significance and no other biomarkers are currently in routine clinical use to help guide treatment planning for these patients. Combined chemotherapy and immunotherapy for advanced and early-stage TNBC has shown a survival benefit for patients with positive PD-L1 expression (e.g., SP142 immune cell > 1%), although response rates in the positive patient group vary widely. Combined assessment of PD-L1 status and TILs density in TNBC has therefore been recommended to help improve patient selection for immunotherapy 30,31 but problems in PD-L1 clinical assays remain 32 . Our study utilises a cell median cut-point definition of PD-L1 positivity and not > 1% immune cell staining, as used in commercial PD-L1 assays such as SP142. Each of the commercially assays (SP142, SP263 both Ventana, 22C3 and 28-8 both Dako) shows variability in detection threshold and variable agreement between clones, which differs significantly from the methods used in the current study, limiting comparisons with studies using these methods 33 . Additionally the epitope target of PD-L1 varies between antibodies, although there is some overlap between the antibody used in our study (E1L3N) and SP142 and SP263 which bind to non-identical epitopes in the cytoplasmic C-terminus domain (22C3 and 28-8 bind to the extracellular domain). 34 Improved characterisation of the stromal PD-L1 immunophenotypes present in TNBC may highlight the importance of the cellular context of PD-L1 expression and potentially also provide data to support the use of immune markers as new routine biomarkers for all patients with TNBC, irrespective of the indications for immunotherapy. PD-L1 (CD274) is a cell surface molecule of the B7 family that may be expressed by immune cells and cancer epithelial cells to inhibit T cell proliferation and induce apoptosis upon binding to its ligand PD-1 35 . Using a retrospective cohort of 244 patients with TNBC, our data demonstrates that high stromal CD68 + PD-L1 + macrophages have incrementally improved prognostic significance to that provided by PD-L1 stromal expression alone or in any other key cellular context, e.g., stromal CD8 + PD-L1 + T cells, carcinoma epithelial PD-L1 + . High PD-L1 + is usually associated with high TILs and a more favourable prognosis [36][37][38] , with a high CD8 + and PD-L1 + population (using monoplex IHC) shown to be prognostically beneficial in an analysis of the IMpassion130 cohort (HR 0.64, 95% CI 0.49-0.83) 39 . However, our data does not find a significant association of outcome with CD8 + PD-L1 + expression and suggests that CD68 + macrophages carry prognostic significance. Prior studies have also identified a role of high CD8 + CD103 + resident memory T-cells (Trm) to be an important predictor of improved outcome in TNBC, but we were unable to address this cellular phenotype in our study due limitations in the number of available targets that could be assessed 40 . The role of macrophages in TNBC was highlighted in a recent single-arm study of 45 TNBC patients receiving neoadjuvant durvalumab and nab-paclitaxel, with high CD68 + PD-L1 + (both stromal and epithelial) associated with improved rates of pCR (73.33% vs. 23.33%, p = 0.053) 41 . Similar findings from the same group have highlighted the improved prognosis which CD68 + PD-L1 + cells carry in non-small cell lung cancer when treated with immunotherapy 42 . Additionally, one other retrospective study of 76 patients in all breast cancer subtypes (23 TNBC cases) found improved rates of pCR with neoadjuvant chemotherapy (74.3% vs. 40%) were associated with high CD68 and PD-L1 expression using monoplex IHC 43 .
Macrophages are typically subclassed within the TIME as classically activated, anti-tumour M1-or alternatively activated, pro-tumour M2-like macrophages, also known as tumour associated macrophages (TAMs). Higher proportions of CD68 + macrophages are seen in TNBC compared to hormone-positive disease, with higher reported unpolarised M0-and M1-like subtype proportions [44][45][46] . However, further studies have shown that TNBC may selectively cause unpolarised macrophages to become TAMs, and thus other studies have shown TAMs to be in greater proportions 47,48 .
The effects of the M2-like phenotype have been more closely linked to cancer progression and metastasis and therefore patient prognosis with SPINK1, LAMC2, IGFBP1, and IL-23A gene expression 47

. A seminal study by
Leek et al. discovered associations between increased macrophage density and angiogenesis, leading to poorer survival outcomes 49 . This implies that macrophages play a pro-tumour role by upregulating endothelial genes, which were later identified as ECSCR, ANGPTL4 and ITGB4 47 . Upregulation of PD-1 and PD-L1 expression is also seen with the presence of TAMs 50 . In the TIME, anti-tumour M1 macrophages promote a cytotoxic response mediated by classic inflammatory cytokines tumour necrosis factor-⍺ (TNF-⍺), interleukin (IL) -1, IL-12, and IL-23 51 . Contrastingly, pro-tumour TAMs recruit more T reg cells and induce apoptosis in CTLs through IL-4 and IL-10 52,53 .
CD163 is a haemoglobin scavenger macrophage receptor that binds to haptoglobin-haemoglobin complexes, and has been used to selectively study M2-like macrophages in TNBC 54 . Recent studies have concluded that high levels of CD163 + TAMs were both associated with poorer rates of pCR in patients receiving neoadjuvant anthracycline-and taxane-based chemotherapy, as well as shorter OS and relapse-free survival (RFS) [55][56][57] . Furthermore, high CD163 + TAMs combined with a low CD4 + , CD8 + , and CD20 + TILs signature was also associated with poorer OS and RFS 58 . These results imply the direction of association is reversed in the M2 macrophage subclass and therefore further study is warranted to substantiate these findings in larger cohorts. CD204, a Class A scavenger receptor associated with angiogenesis, immunosuppression and further tumour proliferation, is www.nature.com/scientificreports/ another TAM marker used in studies investigating invasive breast cancer, with results again suggesting that high macrophage counts were also associated with poorer prognosis [59][60][61] . However, none have looked specifically at TNBC. The timing of chemotherapy (i.e. adjuvant or neoadjuvant) seems to have no effect on the prognostic significance, increasing the value macrophages may have as a clinical biomarker. Single cell RNA-sequencing has found smaller proportions of TAMs within the TIME of TNBC, however high expression of the M2 subtype is associated with poorer overall survival (p = 0.002) 46 . An analysis of metastatic TNBC found non-response to neoadjuvant nab-paclitaxel and pembrolizumab resulted in a lower CD68 signature with higher CSF1R-expressing TAMs 62 . In contrast, treatment response was associated with M1 subtype prevalence expressing CXCL9, CXCL10, and HLA-DR. Given the critical role macrophages have in the TIME, two studies have investigated their effectiveness as targets of immune checkpoint inhibition in TNBC. Preliminary results from a phase II trial have shown that lacnotuzumab (MCS110) in combination with carboplatin/gemcitabine currently provides little benefit for patients with advanced TNBC whilst cabiralizumab will be combined with nivolumab and neoadjuvant carboplatin/paclitaxel for stage II or III TNBC in a trial that is still currently recruiting 63,64 . Both drugs target colony stimulating factor 1 (CSF1). The rationale behind TAM-targeted immunotherapy has six potential mechanisms: suppressing macrophage recruitment, accelerating macrophage apoptosis, inhibiting pro-tumour activities, repolarisation back to M1-like phenotypes, aiding cancer cell phagocytosis, and chimeric antigen receptor macrophage (CAR-M) development 65 . It is yet to be seen which one of these mechanisms will be most effective in improving patient clinical outcomes.
The adoption of mIF and digital pathology brings unique benefits and challenges to solid tumour analysis. Biomarker quantitation using artificial intelligence-driven image analysis is now a reality and will develop into clinical algorithms once standardised and appropriately validated in the near future. The Vectra Polaris™ platform used in this study allows for up to eight targets to be simultaneously visualised on a single specimen, facilitating quantitative spatial analysis and immunophenotype identification 66,67 . It is the foremost used mIF system and has been successfully applied to various solid tumours for research purposes 25,26,66,68,69 . Newer systems, such as Akoya CODEX™, can accommodate up to 40 markers and studies have begun to use this system 67,70 . However, greater standardisation and automation of staining and imaging protocols will further validate the clinical applicability of mIF, including the choice of assay. The potential overlapping of wavelengths associated with each fluorophore can also interfere with spectra detection, with cell detection potentially being problematic and incorrect cell classifications being made. Farkas et al. found a panel utilising greater than seven markers could be problematic and therefore a need for studies of carefully chosen immunofluorescence targets will be required before clinical adoption can occur to avoid crossover and optimise image analysis 24,71 . The selection of regions of interest (ROI) to be analysed is also important to analyse a truly representative sample of the tumour, avoiding hotspots or areas devoid of any immune cell activity. This is one short coming of our study, which uses TMA cores and not whole section images. The variability in immune cell density amongst selected ROIs may be greater, depending on tumour heterogeneity. A recent study comparing mIF to H&E stromal TILs and the SP142 PD-L1 IHC assay found 15 high power fields of the tumour were required to optimise accuracy 72 . Additionally, spatial relationships such as the distance between neighbouring immune cells and the tumour interface, will likely yield further prognostic information but the vast quantities of data will require interrogation using advanced deep learning artificial intelligence algorithms to define their clinical significance 73 . Whilst such algorithms will likely develop and reach clinical application as the uptake of digital pathology gathers pace, until then simplified panels of immune cells will continue to be used for clinical reporting by Pathologists. Our data provides evidence to suggest that a multiplexed panel of CD68 and PD-L1 could potentially improve upon monoplex PD-L1 assays as a general prognostic marker in TNBC. Studies involving cohorts from immunotherapy trials will be required to determine if this may also provide improved predictive value over PD-L1 expression alone to guide patient selection for immune checkpoint therapy.
Differences in methodology in PD-L1 cut-point assessment between studies from clinical trials (e.g. > 1% immune cell staining) and those using cell count /density measurement also limit the comparisons which can be made between such studies.

Conclusion
High dual stained stromal CD68 + PD-L1 + macrophages identifies a subgroup of TNBC patients associated with improved prognosis and incrementally improved predictive value over PD-L1 + alone for BCSS, in multivariate models accounting for age, size, LN status and chemotherapy status. This prognostic immunophenotype provides useful data to further stratify TNBC outcome and aid in decision making for patients under consideration for standard of care chemotherapy.

Methods
Tissue microarray construction. Tissue microarrays (TMAs) were constructed using a Beecher Manual Arrayer MTA-1 (Beecher Instruments, Inc., Sun Prairie, WI, USA), with appropriate areas sampled from the periphery of the tumour block marked up by a breast Pathologist on a H&E slide. Furthermore, 3 cores of normal spleen, 1 core of normal kidney, and normal breast tissue from 1 patient, were included in each of the 9 TMAs as controls. Paraffin sections were cut at 4 μm onto Superfrost™ glass slides (ThermoFisher Scientific, Waltham, MA, USA) and stained for H&E using a Leica automated staining machine (Leica Biosystems, Wetzlar, Germany) in the Department of Anatomical Pathology, NSW Health Pathology, St George Hospital, Kogarah, Australia. States of America), using the following steps: Xylene 2 × 5 min, 100% Ethanol 3 × 1 min, 70% Ethanol 1 × 1 min and distilled water 1 × 1 min followed by 10 min wash in distilled water. The staining conditions for all antibodies were first tested using chromogenic 3,3′-diaminobenzidene (DAB) detection (BOND Polymer Refine detection, Leica Biosystems. #DS9800). Initial antigen retrieving step was performed in Decloaking Chamber™ NxGen (Biocare Medical, Pancheco, CA, United States of America) in citrate followed by EDTA based antigen retrieving buffers (DAKO, AR6 #K8005 and AR9 #K8004) in 110 °C for 5 min. Staining was completed on Leica Bond RX automated immunostainer (Leica Biosystems, Australia). Triple negative breast cancer tissue was used for all optimisation steps. Localisation of IHC staining signal and quality was used as a baseline for comparison for mIF staining.
The TSA-based Opal 9 multiplexing technology was used for immunofluorescence staining (Opal 7-Color Automation IHC Kit, # NEL821001KT; Opal Polaris 480 reagent pack, # FP1500001KT and Opal Polaris 780 reagent pack # FP1501001KT: Akoya Biosciences, Marlborough, MA, USA). Primary antibody conditions determined in the initial DAB optimisation step were applied to the Opal monoplex and multiplex optimisation.
Each biomarker antibody was paired with an individual Opal fluorophore ( Table 5). Pairing of antibody-Opal fluorophore was based on the biomarker co-expression in the tissue and their expected levels of protein expression. Biomarkers expressed in the same compartment were paired with spectrally distanced Opal fluorophores. The Opal fluorophores were used in 1:150 dilution except for Opal Polaris 780 with TSA-DIG 1:150 and Opal Polaris 780 1:25. DAPI was used as a nuclear counterstain. All staining was performed on a Leica Bond RX autostainer (Leica Biosystems, Australia).
In the process of achieving Opal monoplex optimisation, each biomarker was assessed for staining quality and intensity. Acquired monoplex images were unmixed and analysed with Akoya's INFORM software version 2.5.1. Due to the Opal signal and antibody concentration variability we aimed to get an optimal normalised count range of 10-20 while signal to noise ratio assessed by measuring the positive signal versus background with ratio of 10:1 deemed as sufficient. On completion of satisfactory monoplex IF, the mIF protocol was performed using all 8 antibodies on each slide of the TMA cohort. Cores of normal spleen and normal breast tissue were used as internal controls in the TMA slides to ensure staining intensity was comparable across all slides. Antibody concentrations were further adjusted in the multiplex round in normalised counts not meeting the criteria. Image analysis. Fluorescent slides were scanned using the Vectra Polaris 3.0 (Akoya Biosciences, Marlborough, MA, United States of America) using 40 × magnification (Plan APO 40 × /NA 0.75, 0.25um/pixel) and auto-estimated exposure times. Whole slide scan was imaged using 5 epi-fluorescent filters (DAPI, Opal 480, Cy3, Cy5 and Opal 780). Individual TMA cores were selected using the TMA array in the Phenochart software for image acquisition and acquired with auto-estimated exposure times for each epi-fluorescent filter. The full Opal 9 acquisition protocol requires use of 7 epi-fluorescent filters (DAPI, Opal 480, FITC, Cy3, Texas Red, Cy5 and Opal 780) imaging at 20 nm spectral bands as designed for the Vectra Polaris. Multiplex auto-fluorescent slide with no primary antibodies was created and scanned using the same exposure times as labelled multiplex slides. Previously created and assessed spectral library for Opal 9 panel and the auto-fluorescent slide were used for unmixing of the MSI core images in INFORM software. Images of individual and combined colour channels for CD8, CD68 and PD-L1 are presented in Supplementary Fig. S3A-E.
Individual TMA images derived from mIF staining were analysed using an open-source digital image analysis software platform QuPath v0.2.3 (https:// qupath. github. io/) 74. Tissue detection and segmentation into stroma and tumour epithelium was created using trainable machine learning algorithms in the pixel classifiers. Pancytokeratin staining was used to guide tumour epithelium and stroma segmentation, supervised by a Pathologist. Cell segmentation was based on DAPI nuclear staining using the inbuild cell detection algorithm. Two different cell detection algorithms were derived, one for tumour and one for stroma. Phenotyping of all biomarkers was created using the latest multiplex analysis approach available in QuPath v0.2.3, by creating object classifiers. For object classification, we utilised the machine learning algorithms available (random forest). Each classifier was Table 5. Multiplexed immunofluorescence antibodies and their targets. C-AR, citrate antigen retrieval at pH 6.0; E-AR, Ethylenediaminetetraacetic acid (EDTA) antigen retrieval at pH 9.0; Pan-CK, pan-cytokeratin; PD-1, programmed cell death protein-1; PD-L1, programmed death-ligand 1. www.nature.com/scientificreports/ thoroughly trained and verified on multiple selected cell measurements. The combined classifier was applied to each TMA core. Cell counts for all targets were provided for stromal and epithelial compartments per TMA core (average core diameter 1.2 mm = 1.13 mm 2 ). The median cell count value was used as the cut-point for all analyses.
Statistics. Preliminary associations between specific immunophenotype combinations and clinicopathological features were first investigated with a Chi-squared test. Univariate and multivariate OS and BCSS analyses were conducted using Cox proportional hazards modelling, with a p < 0.05 considered significant. A backward selection method was applied to find the most appropriate multivariate models by elimination of redundant variables. Survival predictions were represented with Kaplan-Meier plots. Statistical analysis was completed using IBM SPSS Statistics 26 (IBM Corp., Armonk, NY, USA).
Ethics. Ethics approval was granted by the South Eastern Sydney Local Health District Human Research Ethics Committee at the Prince of Wales Hospital, Sydney (Boost: HREC 96/16 and TNBC: HREC 2018/ETH00138) who granted a waiver of consent to perform research analyses on the tissue blocks. All methods were performed in accordance with the relevant institutional guidelines and regulations.

Data availability
The data is not publicly available due to ethics restrictions but may be accessible on reasonable request to the corresponding author.