Spatial Distribution of Immune Cells Drives Resistance to Neoadjuvant Chemotherapy in Triple-Negative Breast Cancer

Triple-negative breast cancer (TNBC) is the most aggressive and hard-to-treat form of breast cancer. It is defined by the lack of expression of therapeutic targets, including estrogen receptors, progesterone receptors, and Her2 (1, 2). Administration of chemotherapy, based on anthracycline and taxanes as neoadjuvant presurgical treatment (NAC), has been the standard of care in localized TNBC management for a long time. NAC helps reduce invasive surgical procedures and improves patient outcomes. However, only 30% to 40% of patients with TNBC have a complete pathological response (pCR) after NAC; the remaining tumors are refractory, showing no response or progressive disease (pNR). Patients who have pNR are at higher risk of poor outcome (3, 4). The combination of chemotherapy with immune-checkpoint inhibitors (ICI) in the neoadjuvant setting has improved these percentages, underlying the crucial involvement of the immune system in TNBC therapeutic response and overall aggressiveness (5–8). Indeed, the presence at diagnosis of higher rates of tumor-infiltrating lymphocytes (TIL) and immune activation in the tumor microenvironment (TME) are recognized as prognostic factors in TNBC, being positively associated with both pCR and improved survival (9–16).

Despite the importance of the immune system in predicting TNBC outcomes, the relationship is not always consistent. Hematoxylin and eosin (H&E) methods for TIL evaluation average the amount of cells across different zones, flattening the heterogeneity of TIL distribution and not considering the potential impact of spatiality organization. Indeed, the functional heterogeneity of the tumor immune microenvironment and the multidimensional interactions of infiltrating immune cells with cancer and stromal cells within this complex network (17) is emerging as a relevant issue and requires sophisticated spatially resolved approaches to be addressed. Composition and activation state of the infiltrating immune cells also contribute to establish outcomes. A recent study reported that, regardless of the overall extent of tumor infiltration, the intrinsic capacity of TILs to recognize and attack cancer cells is variable, and limited to approximately 10% of intratumoral CD8⁺ cells. These suggest that the majority of cancer TILs are quiescent instead of functional (18). Besides, spatial organization of TILs within the lesion seems to have a prognostic value in TNBC, with the close proximity between TILs and tumor cells positively correlated with pCR in TNBC after NAC (19). Aware of such functional heterogeneity, breast pathologists distinguished TILs into two categories, stromal or intratumoral, based on their spatial distribution and intimate connection with tumor cells. Intratumoral TILs (iTIL) are defined as scattered lymphocytes within tumor nests establishing direct interactions with carcinoma cells, whereas stromal TILs (sTIL) are localized within the stroma, with no direct contact with tumor cells (10). A different relevance of these two categories in terms of cancer immune surveillance and therapy response has been hypothesized, but never proved. The inherent difficulties in visualizing and scoring in a reproducible manner iTILs by H&E staining is the major technical issue that has prevented the implementation of iTIL scoring.

The complexity of TNBC and its associated immune microenvironment are emerging as a 3D structured entity with multiple layers of organization and interactions. This requires new investigation paradigms and spatially-defined approaches to be fully mastered. Here, we have used morphology-guided deep profiling to dissect the multidimensional organization of the TNBC ecosystem at diagnosis and to explore the relative contribution of each cellular component to NAC response. We found that the intimate spatial infiltration of iTILs within the tumor nests was a distinctive feature of TNBC showing pCR. Higher immune burden within the tumor component led to a marked activation of innate and adaptive immune responses and served as a favoring milieu for NAC cytotoxicity. Conversely, pNR lesions were found to be characterized by more efficient angiogenesis and high oxygen-based metabolism that likely ensures energy supply to foster proliferation and structural rearrangement required for tumor progression. In line with what has been already reported (18, 19), we observed that the majority of sTILs did not establish a physical connection with tumor cells but were in a stand-by functional state that did not seem to actively contribute to NAC response.

The study was conducted on a single center retrospective cohort of patients with TNBC, in accordance with the Declaration of Helsinki and approved by the local ethical commitee (protocol number 2020/0011853). Written informed consent was obtained from all living patients.

Eighty consecutive TNBC cases were identified among the cohort of all breast cancers treated with NAC at Azienda Unita Sanitaria Locale-IRCCS of Reggio Emilia (Reggio Emilia, Italy) from January 2013 to December 2019. Histological material from Tru-Cut diagnostic biopsies (pre-NAC) was available for all patients.

Response to NAC was established by two independent pathologists upon evaluation of surgical samples according to current guidelines (20).

We included in the study only patients who were classified into the following two categories that reflect a divergent NAC response: Patients with pCR, and patients pNR. pCR was defined as the absence of residual invasive disease in the breast plus the absence of measurable disease in any axillary lymph node (ypT0/isypN0ypM0). pNR included patients with no evidence of therapy response or partial response to therapy with >50% of tumor cellularity remaining evident, consistent with Pinder and colleagues (21).

Of the initial cohort of 80 patients with TNBC, 28 had pCR. Among the remaining cases, 27 were classified as pNR. A final selection of samples from 24 patients from this cohort was made after quality and quantity assessment of the available bioptic tissue to match GeoMX-DSP requirements and guarantee the best performance of the analytical procedure. The final cohort included 12 patients that had pCR and 12 pNR after NAC. Validation analysis was conducted on a separate cohort of an additional 10 patients with TNBC, 5 had pCR and 5 pNR. Clinical data are summarized in Table 1 (Training set) and Table 2 (Validation set).

The percentage of TILs was evaluated according to Salgado guidelines (10). The number of tertiary lymphoid structures (TLS) was assessed by the pathologists in tumor-adjacent tissues (within 1 mm from tumor burden), including carcinoma in situ components. Lymphoid aggregations with vessels that exhibited high endothelial venule (HEV)-like characteristics (plump and cuboidal endothelial cells) with or without a germinal center were considered TLSs, as previously reported (22). For each patient, the number of TLSs was normalized on the number of tumor chunks available within the diagnostic biopsy. The residual cancer burden (RCB) score was calculated by applying the MD Anderson Cancer Center RCB Calculator (https://www3.mdanderson.org/app/medcalc/index.cfm?pagename=jsconvert3), as reported in Tables 1 and 2.

Tru-cut biopsies (2–4 chunks for each patient) were collected using a 14 to 16 gauge needle and following standard procedures (23). For each Tru-cut biopsy, digital scanning of the whole section used for the analysis was stained with H&E, acquired by Aperio GT450 system (Leica), and is provided in Supplementary Fig. S1. For H&E staining 5-mm formalin-fixed, paraffin-embedded slides were stained with H&E in an automated stainer Leica ST 5020 (Leica Biosystems) using the ST Infinity H&E Staining System (3801698; Leica Biosysthem).

Spatial transcriptomics was performed using the GeoMx digital spatial profiler (DSP; Nanostring Technologies; ref. 24), starting from slides of 5-μm FFPE diagnostic biopsies (≤8 years old). The slides were hybridized with the GeoMx Cancer Transcriptome Atlas (GMX-RNA-NGS-CTA-4; Nanostring Technologies; ref. 25) panel according to manufacturer's protocol. This panel includes 1,800 genes for comprehensive profiling of tumor biology and the TME, focusing on the immune response. Slides were stained with GeoMx Solid Tumor TME Morphology Kit (GMX-RNA-MORPH-HST-12; Nanostring Technologies) according to the manufacturer's protocol. Regions of interest (ROI) representative of the TME were designated by the pathologist (Supplementary Table S1). The median area of the selected ROIs was 54,094 μm² (IQR: 43,986–73,867 μm²). Pan-CK was used to segment each ROI defining specific areas of illumination (AOI) and distinguishing the tumor (CK⁺) from the stromal (CK^–) component. Stromal areas were further stratified using the pan-immune marker CD45. Overall, three classes of AOIs were analysed within each ROI (CK⁺SYTO13⁺, CK^–SYTO3⁺, CK^–SYTO13⁺CD45⁺). For each AOI, probes were collected by the machine and libraries were constructed by GeoMx Seq Code Pack (GMX-NGS-SEQ-[XX]; Nanostring Technologies). Then, AOIs were pooled according to their dimension, purified by AMPure XP beads (A63880 Beckman Coulter) clean up, and resuspended in a volume of elution buffer proportional to the number of pooled AOIs. Libraries were assessed using an Agilent Bioanalyzer, then diluted to 1.6 pmol/L and sequenced (paired-end 2×27) on Illumina NextSeq500, with a coverage of 30 reads for μm². FastQ files were uploaded to BaseSpace Illumina and converted into DCC files by GeoMx next-generation sequencing (NGS) Pipeline App to be associated with the corresponding AOIs.

Data analysis was conducted using GeoMx DSP device analysis suite (26, 27). All samples were checked for quality controls (QC). First, QC was conducted on AOIs to verify the quality of sequencing (>40% sequencing saturation threshold), the number of nuclei collected (>200) and the background effect (negative probe count geomean threshold >6). AOIs not reaching any one of those thresholds were excluded. Then QC was conducted on probes to exclude those outliers for each gene. After quality check, to avoid bias due to the specific molecular features of the different components, we normalized ROIs counts by grouping the regions according to the tissue compartment they belonged to. Thus, we conducted separate normalization for CK⁺ and CK^– AOIs. For each group of segments, we filtered out AOIs that expressed less than 10% of genes. Then we applied a filter on genes by excluding those expressed under the limit of quantitation (LOQ; defined as 2 SDs above the negative probes background value) in less than 5% of the AOIs. For samples that passed filters, raw gene counts were normalized on the third quartile (Q3) of all target genes, since this method was indicated by the manufacturer's procedures as the best to overcome the technical issues related to the differences in data distribution. To do this, we first divided all the genes per AOI by their respective Q3 count, and second we multiplied all the genes in all AOIs by a constant defined as the geometric mean of Q3 counts for all AOIs. After completion of the normalization process, we compared the expression profiles of AOIs derived from pNR with the ones obtained from pCR, in both CK⁺ and CK^– groups. We applied a ratio analysis and to define the list of differentially expressed genes (DEG) we calculated pValue as unpaired t test corrected for multiple comparisons with Benjamini–Hochberg method. Internal resampling (n = 1,000) of original data was conducted to assess the validity of the gene expression comparison analyses. Bioinformatic analyses on gene expression profiles were conducted by R Software v4.1.3 using the following R packages: ggplot2, ggbiplot (function prcomp), topGO, and circlize.

Cellular deconvolution on the basis of normalized gene expression profiles was conducted by GeoMx DSP using “SpatialDecon” pipeline downloaded from GeoScript Hub (https://nanostring.com/products/geomx-digital-spatial-profiler/geoscript-hub/). Scaled abundance score of each cell type was compared between pNR and pCR and pValue was calculated as a two-tailed Student t test.

For data validation, DSP analysis by GeoMX DSP (Nanostring technologies) was conducted starting from slides of 5 μm FFPE tissues. The slides were hybridized with the Immune Pathways Panel Human RNA (GMX-RNA-NCT-HIP-12, Nanostring Technologies) including 84 genes mainly important in the tumor immune microenvironment. The validation analysis was focused selectively on the CK⁺ AOIs. After ROI collection, we performed HyB Code Hybridization at 65°C for at least 16 hours to associate each gene probe included in the initial panel to a specific molecular barcode. Each hybridization reaction was purified and spotted on the cartridge using the nCounter Prep Station (Nanostring Technologies). Finally, molecular barcode detection was performed by nCounter Digital Analyzer (Nanostring Technologies) and raw counts were uploaded on the GeoMX DSP platform to be associated with the corresponding AOIs. For data analysis, first, segments were checked for imaging quality controls and AOIs that did not reach the standard quality were excluded. Then, the signal-to-noise ratio was calculated by dividing each gene count by the mean count of negative controls plus 2 SDs and we defined the list of the expressed genes by considering signal to background ratio >1. Normalization was conducted on the five housekeeping genes available in the panel (OAZ1, POLR2A, RAB7A, SDHA, UBB) by multiplying the raw counts of each expressed gene for a correction factor calculated for each sample as the ratio between the quadratic mean of the housekeeping genes across all the AOIs and the expression value of the gene in each sample. After completion of the normalization process, the expression profiles of tumor AOIs were compared between pNR and pCR by a ratio analysis and pValue as unpaired Student t test was calculated.

IHC was performed using standard protocols and commercially available antibodies in an automated immunostainer (Ventana BenchMark); 3, 3′-diaminobenzidine was used as the chromogen and Harris's hematoxylin as the counterstain (ChromoMap DAB Kit, 760–159, Ventana BenchMark). Primary antibodies were rabbit monoclonal anti-CD3 (2GV6; RRID:AB_2335978), anti-CD4 (SP35; RRID:AB_2335982), anti-CD8 (SP57; RRID:AB_2335985), and mouse monoclonal anti-CD20 (L26; RRID:AB_2335956), all from Ventana BenchMark. An additional primary antibody, anti-ERG (EPR3864; RRID:AB_2098414), was obtained from Novus. OmniMap anti-Mouse HRP (760–4310) and OmniMap anti-Rabbit HRP (760–4311) from (Ventana BenchMark) were used as secondary antibodies.

The datasets generated and analyzed during the current study are available at ArrayExpress (accession no. E-MTAB-13069). All other data generated in this study are available within the article and its Supplementary Data Files or upon request from the corresponding author.

The study design is summarized in Fig. 1A. Twenty-four specimens from 24 patients diagnosed with TNBC that underwent NAC were analyzed (Table 1). This series included 12 patients that had pCR and 12 that had pNR after therapy. In agreement with the literature, the percentage of TILs evaluated by H&E staining was significantly higher (P = 0.03) in the pCR versus pNR groups (Fig. 1B–D). Digital scanning of the H&E stainings of all the Tru-Cut biopsies included in the study is provided in Supplementary Fig. S1. Spatially resolved gene expression was conducted on presurgical diagnostic biopsies selecting CK⁺ tumor and CK^– stromal AOIs. CD45⁺ areas within the CK^– stroma AOIs were also collected (Fig. 1E and F). Principal component analysis (PCA) showed that gene expression profiling efficiently segregated CK^– and CK⁺ AOIs confirming the accuracy of the spatial selection (Fig. 1G). We preliminary explored the spatial distribution of CD20 (Supplementary Fig. S2A–S2C) and CD3 (Supplementary Fig. S2D–S2F), markers of B and T cells respectively, in both CK^– and CK⁺ AOIs of pCR and pNR samples (Supplementary Fig. S2C and S2F). We detected CD20 and CD3 expression in both CK^– and CK⁺ AOIs, indicating their presence not just within the stroma, but also within the intra-epithelial compartment of the lesion. Still, we did not observe a significant difference in the overall expression of these markers in the CK^– stroma component in both pCR and pNR (Supplementary Fig. S2F), whereas a trend indicating their decreased expression was observed in the CK⁺ component of pNR patients, suggesting a potential topological effect in the organization of the immune infiltrate (Supplementary Fig. S2C). We also investigated the expression of some immune checkpoint–associated molecules to monitor potential differences in the distribution of immune-inhibitory signals among the two study groups. The expression of PD-L1 and PD-L2 in the CK⁺ AOIs, as well as PD-1 in the CK^–CD45⁺ AOIs, did not display significant differences among the two groups (Fig. 1H–J). PD-L1 was not expressed in the CK^–CD45⁺ compartment.

A differential analysis based on expression profiles was conducted to establish potential features associated with different responses to NAC and whether these could be preferentially attributed to a specific component. 56.5% (n = 1,017), 55.6% (n = 1,002), and 54.5% (n = 981) of the analyzed genes were expressed in the CK^–, CK^–CD45⁺, and CK⁺ AOIs, respectively (Fig. 2A). The percentage of DEGs between pNR and pCR groups was variable across the three components but the highest transcriptional perturbation was found in the tumor one (CK⁺ AOI 29.1%, CK^– AOI 18.9%, and CK^–CD45⁺ AOIs 12.2%, P = 0.01; Fig. 2B).

Focusing on stroma, we observed that the gene expression profile from CK^– AOIs could partially segregate pCR and pNR samples (Fig. 2C). A total of 192 DEGs were reported, with upregulation of 69 in pCR and 124 in pNR samples (Fig. 2D). Internal validation by resampling analysis (N = 1,000) confirmed the significance of the difference observed between pCR and pNR gene expression profiles (Supplementary Fig. S3A and S3B). Gene Ontology (GO) analysis on these lists identified genes enriched in Immune system process, Response to stress, and Regulation of Metabolic process in pCR samples and in Cell Communication, Signal Transduction, and Regulation of Metabolic process in pNR samples (Fig. 2E). Protein–protein interaction analysis within these categories, identified three nodes among the genes upregulated in pCR stromal AOIs (Supplementary Fig. S3C) corresponding to Response to IFN I and II, Chromatin remodeling and Catabolic processes. In particular, the genes encoding the receptors for IFNA and IFNG (IFNAR1 and IFNGR2) as well as the genes encoding three components of the IFN-activated Transcriptional Complex ISGF3 (STAT1, STAT2, and IRF9; ref. 28) were found upregulated in the stromal AOIs of pCR lesions (Supplementary Fig. S3D), in line with the reported role of IFN signaling in supporting response to anticancer drugs (19).

To analyze the data more deeply, we specifically considered the CD45⁺ cells among the CK^– AOIs. Only marginal differences were observed between the pCR and pNR lesions (Supplementary Fig. S3E). A total of 123 DEGs were found comparing pCR and pNR, 38 of which were upregulated in pCR and 85 in pNR samples (Supplementary Fig. S3F). Response to stimuli and activation as a result of stress were found as enriched biological pathways in both genes lists (Supplementary Fig. S3G). 61% of the 123 genes (75 of 123) were in common with the ones identified in the stromal AOIs, consistent with the fact that immune infiltrate is the predominant component in TNBC stroma. Among these genes, we identified STAT1 and IFNGR2, expression of which was found upregulated also in the CD45⁺ component of stromal AOIs.

Next, we analyzed the tumor component of the samples. PCA analysis showed that gene expression profiles collected from CK⁺ AOIs could efficiently segregate pCR from pNR samples (Fig. 2F). A total of 286 DEGs were identified between the two groups. Of these, 129 were upregulated in pNR samples and 157 in pCR samples (Fig. 2G). An internal validation by means of resampling analysis (N = 1,000) confirmed the strength of this result (Supplementary Fig. S4A and S4B). GO analysis showed that genes upregulated in pNR clustered in four major functional nodes: regulation of angiogenesis and vascular development, oxidative metabolism, transcriptional regulation, and cell–cell adhesion (Fig. 2H and 3A). The top-scoring genes in this analysis were major drivers of tumor angiogenesis vascular development such as VEGFA, VEGFB, ANGPT1, and NOTCH1 and genes involved in oxidative respiration (Fig. 3A–C). In addition, the genes encoding the main transcription factors (TF) known to promote angiogenesis through their induction, such as SOX9, KLF4, HES1, ID2, and ETS2, were found in this analysis. In contrast, genes whose expression was positively associated with pCR converged on immune-response regulation (Fig. 3D). Genes partaking to both innate and adaptive immune-response were found upregulated in pCR versus pNR samples (Fig. 3E–F). For example, genes encoding factors mediating IFNA and IFNG signaling were found enriched in this compartment, including those encoding the IFNA receptor (IFNAR1, IFNAR2) and all four members of the 2′,5′-oligoadenylate synthetase (OAS) family. STAT1, STAT2, and IRF9 (member of the IFNA-induced transcription complex ISG3) were also found upregulated in the pCR CK⁺ AOIs (Fig. 3E; Supplementary Fig. S4C). Genes encoding several pro-inflammatory cytokines (including CXCL9), as well as CD68, a marker of macrophages, were also found as part of the immune-activation signature of the pCR samples. Finally, the upregulation of CD3, CD4, CD5, CD48, and other markers of effector cells suggest a higher infiltration of these cells within pCR tumor areas compared with pNR ones (Fig. 3F). In line with these data, we observed that HLA class I molecules involved in self recognition by the immune system showed higher expression in pCR samples in both CK⁺ (Fig. 3G) and CK^– components (Supplementary Fig. S4D), suggesting a more efficient tumor antigen presentation in pCR patients. In contrast, genes encoding functional markers like CD69, perforin, and granzymes were low expressed and showed no differences between the two groups (Supplementary Fig. S4E and S4F).

We observed that cMYC and genes activated under low oxygen conditions were induced in pCR lesions (29), likely as a complementary phenotype to the active-angiogenetic molecular asset observed in the pNR lesions, suggesting differences in vascular organization and oxygen availability between the two groups (Fig. 3H).

Given the prognostic relevance of the RCB score, we explored its potential correlation with gene expression. Unsupervised clustering showed that the gene expression profile of the CK⁺ AOIs did not stratify patients based on RCB score (Supplementary Fig. S5A). Differential analysis between RCB2 and RCB3 identified 266 DEGs (Supplementary Fig. S5B and S5C). GO analysis showed that many genes significantly upregulated in RCB2 were enriched in antigen presentation processes (Supplementary Fig. S5D and S5E).

We next applied a deconvolution approach to infer cell populations associated with pCR and pNR based on gene expression (Fig. 4A and B). A marked difference was observed in the composition and activation status of iTILs within the CK⁺ AOIs in relation to NAC response (Fig. 4A). Intratumoral immune populations showed a statistically significant reduction in immune effector cells in pNR as compared with pCR lesions. In particular, a significant decrease in the T-cell compartment, including CD4⁺ and CD8⁺ T cells, was observed in pNR samples. In addition, plasma cells and naïve B cells, which were well represented in pCR samples, were significantly reduced in pNR ones. A significant decrease of mature dendritic cells (mDC) was also observed in the pNR group compared with the pCR samples, while mast cells significantly increased in pNR lesions (Fig. 4A; Supplementary Fig. S6A and S6B). This is of potential interest as mast cells have been shown to support cancer aggressiveness through the secretion of growth factors and pro-angiogenic signals and increased accumulation of this population within tumor environments has been correlated with poor prognosis, increased metastasis, and reduced survival in several types of cancer (30).

Conversely, minimal differences were observed in the sTILs within the CK^–CD45⁺ AOIs between pCR and pNR groups (Fig. 4B). No difference was detected in the T-cell compartment, whereas a slight but not significant decrease in the overall B-cell compartment was observed in the pNR group. pNR lesions were characterized by a significant decrease of the natural killer (NK)-cell population and by an increase in the presence of neutrophilis and plasmocytoid dendritic cells (pDC; Fig. 4B; Supplementary Fig. S6C).

To confirm the predicted difference in TIL spatial distribution, we performed IHC analysis of several TIL markers comparing pCR and pNR specimens that were scored with similar percentage of sTILs at diagnosis (Fig. 4C and D; Supplementary Fig. S6F–S6G). An increased percentage of CD3⁺ iTILs was detected in pCR compared with pNR (Fig. 4C). The majority of iTILs were cytotoxic CD8⁺ T cells (Fig. 4D) and, only in a less extent, CD4⁺ helper T cells (Supplementary Fig. S6G). B cells were mainly found in the stromal compartment and more rarely detected in the intra-epithelial part of the lesions (Supplementary Fig. S6F). Together, these results confirmed that tumor immune–infiltrate functionally differs with proximity and physical contact with cancer cells and that a hot immune TME characterized by a prominent effector population is required for an efficient response to NAC.

We also explored potential alterations in the organization of the non-immune TME. A deconvolution analysis was applied to analyze the distribution of endothelial cells and fibroblasts within both tumor and stroma AOIs (Fig. 4E–F; Supplementary Fig. S6D–S6E). No significant differences were observed in these two populations in either compartment. However, a nonsignificant trend of increased endothelial cells cells in CK^– stromal regions was observed in the pNR lesions as compared with pCR. This is in agreement with the pro-angiogenetic signals that we observed in the expression analysis (Figs. 2H and 3A–B). Conversely, fibroblast populations seemed to increase in the CK⁺ AOIs of pNR lesions, even if this difference was not significant. In Fig. 4G–H, the spatial distribution of VEGFA and its receptor FLT1 expression counts is shown in representative pCR and pNR lesions respectively. Besides confirming that pNR lesions display a higher expression of both factors (especially VEGFA) as compared with the pCR samples, this analysis showed that VEGFA expression was primarily localized at the level of CK⁺ AOIs, indicating that tumor cells were the primary sources of the pro-angiogenetic stimuli that support pNR lesions. IHC staining confirmed an increased density of endothelial cell whereas no particular differences were observed in the vasculature organization and morphology between pCR and pNR lesions (Supplementary Fig. S6H).

To validate the observations using data generated from the training set, we investigated an independent cohort of 10 TNBC (5 pCR, 5 pNR; Supplementary Fig. S7A) using the GeoMx Immune Pathways Panel and including a selected number of genes associated with TME (Fig. 5A). Despite the limited number of genes investigated, we confirmed that an immune-activated condition was associated with pCR lesions. This was demonstrated by the fact that markers of immune activation were more expressed in the CK⁺ area of pCR samples as compared with the CK⁺ area of pNR samples, whereas immune-inhibitory mediators tended to be overrepresented in pNR lesions (Fig. 5B). Even if the limited nature of this cohort prevented significant conclusions, unsupervised clustering suggested that an immune activation signature tended to stratify the two subgroups better than the immune inhibitory one, supporting the idea that a hot TME is a major driver of improved response to NAC in this disease (Supplementary Fig. S7B and S7C).

TNBC can be an immunogenic disease and the degree of immune activation deeply affects NAC response (11–13). However, although these data opened the way to the addition of ICIs to the neoadjuvant setting in early TNBC (6), the organization of TNBC immune activation and how it is established remain to be fully elucidated.

The high spatial complexity of TNBC, its microenvironment, and the reciprocal interplay between cancer cells and other cells in the TME represent major limitations to the possibility of resolving the molecular dynamics that underline NAC resistance. The assessment of the overall extent of lymphocyte infiltration alone is not sufficient to explain clinical outcomes, but the topology and the physical relationships among the cell types within the tumor milieu count as determinant factors in tumor response to therapy (17, 18). A bulk transcriptomics approach does not allow to take into account such complexity, resulting in a flat representation of the system in which it is hard to reconstruct significant differences among samples.

Here we used a spatial transcriptomics approach to dissect the principal components of TNBC lesions and analyze potential differences in a retrospective cohort of patients with TNBC that displayed divergent responses to NAC. Even if preliminary, our data define a model in which different responses to NAC are determined by an innate circuit of interactions between cancer cells and the microenvironment (Fig. 5C). In particular, an efficient response to therapy is associated with a high degree of spatial contamination and physical contacts between cancer and immune cells. This is supported by a sustained activation of both innate and adaptive immune responses. As previously reported, intratumoral recruitment, spatial organization and activation of B cells and CD8⁺ T cells seemed to be the critical foundation for effective therapy response and patient outcome (18). Likely, the higher presence and the spatial organization of immune effector cells within the tumor component act to create a favorable environment synergizing with cytotoxic therapy and improving response.

Pathologists score TIL percentage applying a morphologic evaluation and assessing the average extension of their infiltration within tumor stroma (10). The overall extent of immune infiltration has been reported to be a key determinant in early TNBC (31). However, this score does not consider other relevant variables, including heterogeneity in the TIL distribution patterns and the presence of TILs in the intra-epithelial compartment of the lesion. These aspects are particularly hard to score and represent technical limitations that can potentially reduce the clinical value of TIL scoring and require new and more sophisticated analytical approaches to be addressed. Spatial transcriptomics and artificial intelligence applied to digital pathology may represent two complementary opportunities to face this challenge and can be integrated with the current evaluation model helping the pathologist to improve patients’ risk-based stratification. In particular, the employment of automated H&E-stained image analysis algorithm and spatial statistics can provide quantitative parameters relative to TIL pattern distribution that could be integrated with the overall TIL infiltration score (32). Spatial transcriptomics may also help increase understanding of how the functional heterogeneity within TIL populations heavily impacts on the overall extent of the antiturmoral immune response.

In this regard, our data show in a quite ineluctable way that iTILs are a functionally distinct population from sTILs and that they represent the major determinant within the TIL fraction in supporting response to NAC in TNBC. This is consistent with previous studies that reported that most of the TILs that infiltrate the lesion and establish a weak physical connection with tumor cells are in a functional tolerant state and do not substantially contribute to therapy response (18). The physical distance measured between CD8⁺ T cells and tumor cells was reduced in immunologic hot tumors compared with cold lesions, further supporting the essential role of spatial distribution in defining the success of immune response. Finally, Fassler and colleagues recently employed machine-learning and computer vision algorithms to show that, in breast cancer, large aggregates of peritumoral and iTILs are associated with longer patient survival, whereas the absence of iTILs is associated with increased risk of recurrence (33).

In line with the hypothesis that spatiality is a major determinant of immune dynamics in TNBC, there is growing interest in better understanding cancer-associated TLSs (34). These structures are ectopic vascularized aggregates of lymphoid cells within tumoral tissue that resemble secondary lymphoid organs, mainly constituted by interacting B and T cells. An increase in TLS presence is associated with a better response to NAC and improved outcome in patients with TNBC (35). However, TLS score in diagnostic biopsies of TNBC is less informative compared with TLS score in surgical specimens or TIL score, probably due to the low numbers of TLS in biopsies, or to the difficulty in correctly scoring them, thus limiting their use in this context (22). Consistent with these data, the number of TLSs in Tru-Cut biopsies in our patient cohort was limited and the difference between pCR and pNR patients did not reach statistical significance (Tables 1 and 2). Still, our data indicate that patients that had TLSs in their diagnostic biopsies preferentially belonged to the pCR group and displayed higher TIL score, thus suggesting a positive correlation between TIL and TLS scores. The TLS score appears to be particularly relevant in predicting ICI efficacy in solid tumors (36), thus future studies applying spatial profiling in TNBC patients treated with pembrolizumab will be required to further address the clinical relevance of TLSs in this context.

Our data indicate that immune stimulation associated with improved NAC response is driven by an extensive hyperactivation of the IFN signaling in both tumor and stromal components. Type I IFNs stimulate innate immunity by enhancing antigen presentation and the release of immune response mediators, as well as adaptive immunity through antibody production by B cells and enhancing the effector functions of T cells (37). This is consistent with the observation that HLA-class I molecule expression was impaired in pNR patients, suggesting less efficient antigen presentation is occurring in these tumors, thus explaining a defective activation of adaptive immunity. Type II IFN (IFNG) is mainly produced by activated NK cells and T cells and acts on macrophages, which become hyper-responsive to inflammatory stimuli (38). Collectively, the data suggest that the activation of the type I and type II IFN axes improves the immunity and inflammation in TNBC tumors, prompting the efficacy of NAC chemotoxicity, as also reported in ER− patients with breast cancer (39).

In addition to confirming a role for type I and type II IFN in TNBC responsiveness to NAC, our data also shed light into the mechanisms that allow IFN-mediated program to be executed in this context. Indeed, genes encoding the IFN-activated transcription factors STAT1, STAT2, and IRF9 were found upregulated in the pCR CK⁺ AOIs and confirmed in the validation analysis. These three factors heterodimerize in response to IFN stimulation to form the ISGF3 transcriptional complex, which translocates to the nucleus and binds to interferon-stimulated response elements (ISRE) in the promoter of IFN-stimulated genes (ISG) and mediating execution of IFN response (40). In light of these results, additional analyses focused on the IFN signaling cascade, using a multiparametric protein evaluation has the potential to shed light on the extent of contribution of this pathway in driving immune response and response to therapy in TNBC.

Other studies have investigated the spatial distribution of immune cells in TNBC applying a proteomic approach to predict patient outcome (41, 42). These researchers found significantly increased levels of cytotoxic proteins such as Granzyme A and B in responders. These findings are apparently in contrast with our data, which did not show a deregulation of Granzymes and other relevant T-cell cytotoxicity–related molecules, which were barely detectable by our transcriptome approach. A possible explanation is that upon activation, tumor-infiltrated immune cells substantially increase production and release of cytotoxic proteins, but only minimally change the corresponding mRNA levels, as previously reported in NK cells (43). Proteomic and transcriptome spatial approaches can provide distinct and complementary information to study TNBC complexity.

We found that pNR lesions were characterized by intrinsic features of aggressiveness (Figs. 2H and 3). Increased structural plasticity, loss of epithelial structures, and cytoskeleton reorganization marked the predicted phenotype of pNR tumors. Transcriptional rewiring and marked pro-angiogenic signaling supported these acquisitions. This observation likely holds multiple implications (44). First, an appropriate oxygen supply guarantees the required energy need to support tumor progression (45). Indeed, genes partaking to oxidative metabolism were enriched in pNR as compared with pCR tumors. In addition, an important regulatory loop between angiogenesis and immune activation in cancer has been documented. Angiogenesis inhibits immune activation against cancer at different levels. VEGF, a master driver of angiogenesis, was reported to interfere with maturation of dendritic cells, thereby suppressing T-cell priming (46). Consistent with this, in our analysis, strong upregulation of VEGFA and VEGFB in pNR samples was associated with a reduction in mDCs and a significant reduction of the T-cell compartment, including CD8⁺ T cells, in the CK⁺ AOIs. Conversely, immune-modulatory signals such as IFN act by blocking the processes of neo-angiogenesis favoring the normalization of vessels in tumor lesions (47). Our data fit within this framework and underline how this functional interplay in the context of TNBC is fundamental in mediating an efficient response to NAC therapy, paving the way to the combination with anti-angiogenetic compounds to further promote an immune-reactive TME.

Our study has an exploratory function and aimed at characterizing the biological mechanisms underlying the response to NAC therapy in TNBC. The cohort that we analyzed consisted of 34 TNBC cases and its size is in line with other studies that apply spatial-profiling approaches (48). However the number of regions that we were able to collect from each patients was limited by the dimension of the Tru-cut biopsies. Consolidation of these results in an independent validation study is mandatory in order to determine whether these findings have important implications in clinic. Another recent study applying spatial proteomics approach to an independent cohort of TNBC patients supports our data and confirms the prognostic importance of intra-epithelial microenvironment, which has a distinct and unique characteristics compared with the stromal compartment (42).

Anticipating response to NAC by the application of predictive markers would allow the a priori identification of patients for whom chemotherapy alone would not be effective, thus helping the decision of the best therapeutic approach. This is particularly relevant in the NAC scenario in which ICIs have been added to NAC protocols (5). In this setting, RCB score has been suggested to accurately predict patient survival and an increased ICI response was detected in RCB2 patients, but not in RCB3 patients (49). Although we did not observe a significant difference among tumor spatial trascriptomes based on RCB scores, we found some DEGs comparing RCB2 to RCB3 patients. Upregulated DEGs in RCB2 patients were mainly involved in the antigen-presentation process, further supporting the previous observations that these patients may benefit more of an immunotherapy approach compared with RCB3 patients. Overall, our data indicate that the development of a robust prediction tool in the neodjuvant setting of TNBC patients cannot be limited to a molecular signature or a TIL evaluation assessed by the pathologist, but must consider other clinical variables, as well as spatial heterogeneity, considering both topology and reciprocal interaction of the different cells within the tumor ecosystem. In this regard, the implementation of machine learning models that integrate clinical, molecular and digital pathology data seems currently the best possibility to overcome these limitations and develop robust predictors for therapy choices (50).

No author disclosures were reported.

B. Donati: Data curation, formal analysis, validation, investigation, methodology, writing–original draft, writing–review and editing. F. Reggiani: Data curation, investigation, writing–original draft, writing–review and editing. F. Torricelli: Data curation, formal analysis, methodology. G. Santandrea: Investigation. T. Rossi: Methodology. A. Bisagni: Investigation. E. Gasparini: Resources, data curation. A. Neri: Supervision, writing–review and editing. L. Cortesi: Writing–review and editing. G. Ferrari: Resources. G. Bisagni: Supervision, writing–review and editing. M. Ragazzi: Conceptualization, data curation, formal analysis, validation, investigation, methodology, writing–original draft, writing–review and editing. A. Ciarrocchi: Conceptualization, data curation, formal analysis, supervision, funding acquisition, visualization, writing–original draft, writing–review and editing.

This work was supported by the Viva Vittoria Reggio Emilia Project through the following patients' organizations: Senonaltro, Vittorio Lodini, Aibat, Andos, il giorno dopo. We wish to thank GRADE onlus for support. F. Reggiani was supported by Fondazione Umberto Veronesi. We also wish to thank Marina Grassi and Elenonora Zanetti for technical support and all the members of the lab for stimulating discussion.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.