WSG-ADAPT-TN (NCT01815242) is a prospective, multicenter, controlled, non-blinded, randomized, investigator-initiated trial in women ≥18 years of age with histologically confirmed unilateral, primary invasive cT1c-cT4c or cN+ TNBC (centrally confirmed hormone receptor positivity in <1% of tumor nuclei and HER2 score of 0–1+ or 2+ with negative fluorescence in situ hybridization test). Histopathology subtype was assessed by local pathologists at the time of diagnosis. pCR (ypT0/is, ypN0), assessed after 12 weeks of NACT, was a primary endpoint of the study. Surgery was performed within 3 weeks after NACT. In case investigators opted for further NACT, prior histological confirmation of non-pCR by core needle biopsy was obligatory. Recommended poststudy therapy (neoadjuvant and/or adjuvant) followed national guidelines. Core biopsies used for the present analysis were performed at the time of diagnosis and after 3 weeks of NACT (online supplemental figure 1).

The PAM50 assay was used to classify TNBC into five intrinsic molecular subtypes (luminal A, luminal B, HER2-enriched, basal-like and normal-like) as described elsewhere.22 The expression signatures of 50 genes included in PAM50 were compared with each of the five intrinsic molecular subtype centroids using Spearman’s rho correlation levels as a distance measure (the subtype was assigned to the nearest centroid).

Ki-67 was assessed by central pathology using rabbit monoclonal anti-Ki-67 antibody 30–9 (Ventana Medical Systems, Tucson, Arizona, USA). Early Ki-67 response in the 3-week biopsy was defined as >30% decrease in Ki-67 versus baseline or <500 invasive tumor cells in the 3-week biopsy as previously described.21

Stromal TILs were evaluated by a two-independent observer approach following the guidelines of the International TILs Working Group 201423 as previously described.12 Briefly, the evaluation was based on H&E staining on digital whole slide image (WSI), and consensus was obtained by joint re-evaluation. In single cases, assessment of a third evaluator or additional information to visualize tumor composition (eg, CK5/14 and E-cadherin immunohistochemistry (IHC) staining) was considered. In case of heterogeneous spatial distribution, results were averaged. Brisk focal infiltrates (‘hot spots’), obviously perivascular infiltrates, and tertiary lymphoid structures were excluded from the analysis. For statistical analysis, TILs levels, recorded as percentage, were expressed as a continuous variable.

Multiplexed IHC was performed on 3 µm-thick sections cut from the formalin-fixed paraffin-embedded tissue blocks from surplus material after central pathology review, which was available for 83 out of 385 study cases. These additional sections were stored at 4°C in the dark to minimize antigen aging over time until further processing. Inclusion criteria for the multiplex study were (1) availability of paired biopsies before and on treatment, (2) sufficient size of the biopsies on visual inspection, (3) representation of cases covering the entire study recruitment time, and (4) sufficient remaining representative tissue for quantitative analysis on microscopic inspection and comparison with the initial H&E staining, resulting in a series of 66 fully evaluable cases.

Assay optimization was performed following the manufacturer’s instructions (OPAL Multiplex IHC Assay Development Guide; Akoya Bioscience, Menlo Park, California, USA). Slides were deparaffinized; initial antigen retrieval was performed by microwave cooking in Tris-buffered saline (TBS) at pH 9; and blocking of unspecific protein binding was performed using Protein Block Serum-free solution (Agilent/Dako, Hamburg, Germany). Subsequent antigen retrieval and deactivation of the preceding staining step were performed according to the previous optimization by microwave cooking either in TBS at pH 9 or citrate buffer at pH 6. Consecutive IHC staining using the OPAL 7-plex fluorescence system was performed using the following primary antibodies: anti-CD4 (clone SP35, Zytomed Systems), anti-CD8 (clone C8/144B, Agilent/Dako), anti-PD1 (clone NAT105, Cell Marque), anti-CK7 (clone OV-TL12/30, Agilent/Dako), anti-CD73 (clone EPR6114, Abcam), anti-PDL1 (clone QR1; Quartett, Berlin, Germany), and a laboratory-developed assay24 25 robustly detecting PDL1 that is not linked to a specific therapy and was validated in several QuIP ring studies (https://www.pdl1portal.eu/). Cell nuclei were stained with 4′,6-diamidino-2-phenylindole. The following fluorophores in combination with the tyramide signal amplification system were used for detection of bound antibodies: Opal 520, Opal 540, Opal 570, Opal 620, Opal 650, or Opal 690. Fluoromount-G mounting medium (Thermo Fisher Scientific, Braunschweig, Germany) was applied to cover slides before imaging.

WSI scanning was done at ×20 magnification using the Vectra Polaris instrument (Akoya Bioscience, online supplemental figure 2). Three-channel fluorescent WSIs were used to select regions of interest (ROIs) for subsequent targeted scanning of image stacks at ×40 magnification across the visible spectrum (420–720 nm) for multispectral imaging. The full available tumor area including the adjacent stromal area was included, excluding from analysis only ROIs containing obvious artifacts, pre-existing mammary gland tissue, and stromal areas with brisk perivascular inflammatory infiltrates outside of the tumor areas. Spectral libraries were generated using single-stained scans of tonsil tissue for each reagent and image deconvolution was performed with the inForm Advanced Image Analysis software (inForm V.2.4.8, Akoya). Training for the machine learning-based segmentation, tissue classification and cell phenotyping algorithms was performed using the inForm software on 20 images covering the full spectrum of expected variability regarding staining intensities and cell densities. Batch analysis and data export into text format comprised 2188 images representing all evaluable tissue areas in the data set. Visual quality control of all analyzed ROIs was performed by reviewing all composite images in the context of the quantitative output using Image Miner software V.2.7.0 (Definiens, Munich, Germany) (FF), and in-depth control of quantitative output in context of the corresponding H&E-stained sections was performed in 10% of randomly selected cases (FF and MC).

Clinical characteristics between patients included in this immune response assessment study and all other patients included in the WSG-ADAPT-TN trial were compared by independent two-sample t-tests, Wilcoxon rank-sum tests, χ² tests or Fisher’s exact tests. Cellular marker levels at baseline and after week 3 were summarized by median values and 25th and 75th percentiles. Changes between the two time points were evaluated with the Wilcoxon signed-rank test.

Correlations between cellular marker levels and TILs levels (evaluated as a continuous variable) were calculated using Pearson’s correlation coefficient (r). Associations between pCR and patient characteristics, as well as baseline cellular marker levels in stroma and tumor, were evaluated with univariable logistic regression. ORs were calculated per IQR, for example, ORs compared two patients with marker values at the 75th and 25th percentiles. Associations between pCR and changes in cellular marker levels between baseline and week 3, dichotomized as decrease versus increase, were evaluated with logistic regression adjusted for baseline marker levels. A trend test was based on the significance of the continuous marker level. Separate analyses were performed for measurements in stroma and tumor. Dynamics in T cells, PDL1 and PD1-positive CD4 and CD8 cells during 3 weeks of therapy were evaluated by grouping tumors into cold and hot (below-median and above-median marker levels at baseline, respectively), according to a recently proposed nomenclature.14 However, these definitions are currently emerging and might change, depending on tumor type and therapy. For analyses of combined stroma and tumor measurements, tumors were categorized into cold, hot and altered (figure 1). Tumors were considered cold when marker levels in stroma and tumor were below the respective median value, hot when marker levels in tumor were above the median value, irrespective of levels in stroma, or altered when marker levels were below-median in tumor and above-median in stroma. We used univariable logistic regression to evaluate the association between pCR and transitions between the different states under therapy, that is, ‘cold-to-hot’ and ‘hot-to-cold’, from or to altered, and ‘no change’. For all test results, two-sided p values were reported and interpreted as significant at 5%. P values were not corrected for multiple testing because of the exploratory nature of this study. To aid interpretation, ORs for pCR below 0.5 or above 2.0 were considered potentially clinically meaningful. Statistical analyses were performed using STATA SE V.6 software.

Between June 2013 and February 2015, 336 patients were randomized at 45 centers in the WSG-ADAPT-TN substudy (figure 2). pCR occurred in 36% of patients, that is, 28.4% of patients on nab-paclitaxel+gemcitabine and 44.5% of patients on nab-paclitaxel+carboplatin (p=0.00421). The 7-plex IHC and multispectral imaging were performed in a subset of sequential biopsies from 66 patients obtained at baseline and after 3 weeks of NACT. The median age of this group was 49 years, and 54.5% (n=36/66) were premenopausal (table 1). Most tumors were of basal type according to PAM50 (86.4%, n=57). Most patients had cT2 (65.2%, n=43) and cN0 (77.3%, n=51) cancer. Forty (60.6%) and 26 (39.4%) patients received nab-paclitaxel+gemcitabine or nab-paclitaxel+carboplatin, respectively. pCR was achieved in 36.4% (n=24) of tumors, and early response rate (Ki-67 decrease or low cellularity) was 43.1% (n=28). Characteristics of the 66 patients with immune response assessment were not statistically different from the characteristics of the remaining 270 patients with the exception of a lower rate of early response (43.1%, n=28/65 vs 58.9%, n=116/197, p=0.026; table 1).

Compared with baseline levels, the absolute number of CD4, CD8 and total T cells significantly increased after 3 weeks of NACT in stroma (median change: 29.1, 20.1 and 49.8 cells/megapixel, respectively) and in tumor (median change: 10.3, 21.6 and 39.0 cells/megapixel, respectively; table 2). NACT also induced significant accumulation of PD1-positive CD8 cells in stroma (median change: 3.5 cells/megapixel) and in tumor (median change: 9.5 cells/megapixel). Furthermore, we observed a significant increase of CD73 cells in stroma (median change: 6.4 cells/megapixel) and PDL1-positive cells in tumor (median change: 32.3 cells/megapixel).

Correlation between the levels of TILs and immune markers either at baseline or their changes during the 3 weeks of NACT was low to moderate (table 3). The highest correlation coefficients (Pearson’s r) were noted for change in T-cell levels in stroma (0.52) and tumor (0.40), baseline PDL1 levels in stroma (0.47) and tumor (0.45), and for change in CD4 cell levels in stroma (0.46).

Baseline levels for any of the cellular markers in tumor or stroma were not associated with pCR (online supplemental table 1). Increases in the baseline CD4 to CD8 ratio in stroma of an IQR appeared associated with about a twofold reduced chance of pCR, although the association was not statistically significant.

A decrease in the PD1-positive CD4 count in tumor was significantly associated with lower pCR rates compared with an increase (OR=0.25, 95% CI 0.08 to 0.80, p-trend=0.023; table 4). The pattern in the stroma was similar, though not statistically significant. Lower pCR rates after a marker decrease (compared with an increase) were also observed for CD8, T cells, PD1 in CD8, and PDL1, although trends were not statistically significant.

For T cells and PDL1 separately in stroma or tumor and both compartments combined, transitions from cold at baseline to hot at week 3 were associated with similar pCR rates as no change (figure 3). Increased ORs appeared to suggest higher pCR rates after cold-to-hot transitions in PD1-positive CD4 (in tumor and stroma separately) and PD1-positive CD8 cells (in tumor and in tumor and stroma combined). Based on PD1-positive CD8 cells, hot-to-cold transitions separately in stroma or tumor and combined were associated with significantly reduced pCR rates compared with no change. However, these results were based on the absence of pCRs in patients with tumors changing from hot to cold. Lower pCR rates in tumors after hot-to-cold transitions compared with no change were also observed for T cells and PD1-positive CD4 cells, but numbers were too small to reach statistical significance. In a combined analysis of changes in tumor and stroma compartments, TNBC with altered T-cell and PD1-positive CD4 cell levels had higher odds of pCR than those with no change (figure 3).