The spatial immune landscape predicts outcome and reveals the central role of tumor-associated macrophages in inflammatory breast cancer biology.

Introduction
Inflammatory breast cancer (IBC) is an uncommon type of locally advanced breast cancer characterized by rapid progression and an unfavorable prognosis. Although only accounting for around 2% of all breast cancers, IBC is believed to be responsible for 7% of breast cancer-related mortality [1]. This can be explained by the aggressive behavior of IBC: almost all patients have nodal disease at the time of diagnosis, around 30% of the patients present with metastatic disease, and the 5-years survival rate is less than 50% despite multimodality treatment [2–4]. Unfortunately, there have been no significant changes in the survival rate of patients with IBC [5], partly due to the limited understanding of IBC biology and the lack of IBC-specific treatment options. Therefore, there is a compelling need for IBC-specific research, especially because IBC can also be considered a human model for fast-spreading and aggressive breast cancer in general.
Biologically, IBC is distinguished from other forms of breast cancer by its highly inflammatory tumor microenvironment, characterized by lymphovascular tumor emboli, elevated cytokine signaling and prominent infiltration of immune and myeloid cells. This suggests that immune–tumor interactions play a pivotal role in its unique clinical behavior [6, 7]. However, it is not clear how infiltrating immune cells can play a role in the development and aggressive phenotype of IBC. Our previous research showed that the immune microenvironment of IBC consists of the same number of stromal tumor-infiltrating lymphocytes (sTIL) as subtype-matched non-inflammatory breast cancer (nIBC) [8], although a recent study reported lower sTIL scores in IBC than those previously reported in the literature [9]. More recent evidence based on CIBERSORT deconvolution data revealed that the composition of these infiltrating immune cells appears to be highly similar between IBC and nIBC, albeit with some differences, most notably in M1 macrophages [10]. However, none of the studies examined the spatial composition of the infiltrate. Furthermore, the infiltrating immune cells might also be in different functional states, caused for example, by the higher expression of PD-L1 [8]. In nIBC, the impact of infiltrating immune cells on prognosis has been extensively studied, and the presence of sTIL is associated with a better outcome and response to neoadjuvant chemotherapy (NACT), especially in more aggressive subtypes [11]. In IBC, the prognostic significance of immune infiltration is less understood, although sTILs may also be linked to better outcomes in IBC [8]. Moreover, a 107-gene signature predictive of response to neoadjuvant chemotherapy in IBC includes multiple genes involved in CD8+ T-cell activation and signaling, underscoring a potential role for adaptive T-cell immunity in mediating anti-tumor effects [12]. In addition, antibody-producing B-cells have also been associated with a favorable prognosis, suggesting a broader contribution of the immune microenvironment to disease outcome in IBC [13]. The function of cytotoxic T-cells and B-cells is controlled by regulatory T-cells (Tregs), which are considered to play a role in local immunosuppression by secreting cytokines such as TGF-β and IL-10 [14]. Cytotoxic T-cells, plasma cells (effector B-cells), and regulatory T-cells are also the three most common types of lymphocytes in the tumor microenvironment (TME) of IBC [10]. Therefore, the detailed analysis of the spatial distribution of these lymphocytes may help to determine how they contribute to the aggressive phenotype and anti-tumor response in IBC.
The most common type of immune cell in the TME of (inflammatory) BC are macrophages [10, 15]. They can have tumor growth-suppressive or -promoting properties, depending on their polarization [16]. Tumor-associated macrophages (TAMs) seem to be more abundant in both the TME and the surrounding normal healthy breast tissue of IBC patients than in nIBC patients [17]. Moreover, in vitro studies suggest that TAMs in IBC secrete cytokines that promote tumor progression and suppress the anti-tumor immune response [18–21]. Finally, in an in vitro co-culture IBC model, treatment with antibodies against M-CSF1, which plays a role in M2 polarization of TAMs, reduced the aggressive behavior of IBC tumor cells, in the absence of mature T- and B-cells [22]. Thus, TAMs seem to play a fundamental role in the TME, and further research on their involvement in the invasive properties of IBC is warranted. In this study, we aimed to elucidate how the anti-tumor immune response is different in IBC compared to molecular subtype-matched non-inflammatory breast cancer (nIBC), how this immune response can contribute to the fulminant course of IBC, and which components of this immune response might have a prognostic role.

Methodology

Patient selection
We previously reported a retrospective cohort of consecutive IBC patients who had their initial diagnosis and complete treatment at ZAS Sint-Augustinus (Antwerp, Belgium), Antwerp University Hospital (Edegem, Belgium), or Institut Paoli-Calmettes (Marseille, France) between June 1, 1996, and December 31, 2017 [8]. All cases (n = 161) had complete hospital records, were pathologically confirmed as invasive carcinoma, and were diagnosed with IBC using the clinical definition agreed upon by international experts [23]. The included samples were pre-treatment primary tumor samples from patients that corresponded to diagnostic biopsies (open or core biopsy). Estrogen receptor (ER) and progesterone receptor (PgR) expression was assessed using validated immunohistochemical tests and was defined as positive if the Allred score was ≥ 3/8. Tumors were considered HER2-positive if the HER2/CEP17 ratio was ≥ 2.0 with an average HER2 copy number ≥ 4.0, or if the ratio was < 2.0 with ≥ 6.0 copies per cell. A pathological complete response (pCR) was defined as the absence of residual invasive carcinoma in the resected breast specimen (mastectomy) and in all sampled regional lymph nodes after the completion of NACT. The NACT regimen was anthracycline/taxane-based, and most HER2 + patients received targeted therapy (n = 43/48). This study was approved by the Ethics Committee of the Antwerp University Hospital (Filenumber: 16/33/338).

Control group
To compare the composition of the immune infiltrate between IBC and nIBC, we included a retrospective cohort of 115 subtype-matched patients with non-IBC breast cancer. This cohort was sampled at random using the cancer registry from ZAS Sint-Augustinus or Institut Paoli-Calmettes in 2006 to match the period during which most IBC cases were diagnosed. All patients received adequate local and systemic treatments after a pathologically confirmed diagnosis. Exclusion criteria included IBC, previous breast cancer treatment, diagnosis of ductal carcinoma in situ, or loss of follow-up.

Stromal tumor infiltrating lymphocytes (sTIL) and PD-L1 scoring
Two researchers evaluated PD-L1 expression and infiltration with sTIL as previously reported [8]. In brief, TIL scoring was performed on hematoxylin and eosin (H&E)-stained 5-µm sections of formalin-fixed paraffin-embedded (FFPE) pre-treatment tumor tissue, according to the recommendations of the International TILs Working Group [24]. PD-L1 expression on infiltrating immune cells was assessed using a validated PD-L1 assay (clone SP142, Ventana Benchmark). A score of 0 ( PD-L1−), 1, 2, or 3 was assigned for < 1%, ≥ 1% but < 5%, ≥ 5% but < 10%, or ≥ 10% PD-L1-positive cells per tumor area, respectively. In cases of discrepant results between the researchers, a consensus score was determined.

Staining and scanning (Fig. 1)
Consecutive (5-µm thick) FFPE slides were stained for the following validated antibodies: CD79α (clone: JCB117, Dako, M7050, 1:400) for activated B-cells and plasma cells, CD8 (clone C8/144B, Dako, M7103, 1:150) for cytotoxic T-cells, FOXP3 (clone: 236 A/E7, eBioscience, 14–447, 1:50) for regulatory T-cells (Tregs), and CD163 (clone: MRQ-26, Ventana, 760–4437, Ready-to-Use dispenser) for Tumor-associated macrophages (TAMs). Staining was performed using Bond/Leica or Benchmark/Ventana Autostainers. Antibodies were visualized using HRP 3,3′-diaminobenzidine (DAB) and the slides were counterstained with hematoxylin. All the stained slides were digitized using a digital slide scanner (3DHISTECH, Hungary; 20x objective, 40x optical magnification at 0.25 μm/pixel.). The first quality check was performed after the staining and scanning. During quality control, we checked for out-of-focus scanning, staining artifacts, and disconnected tissue. If possible, the slides were scanned again; otherwise, patients were excluded.

Image analysis (Fig. 1)
For robust and verifiable immunohistochemical quantification, we used Visiopharm software.
Standard H&E-stained sections of FFPE tumor tissues were used to mark the region of interest (ROI) on the excisional or trucut biopsy sections (annotation). A more diffuse and less expansive growth pattern can be a characteristic of IBC and might complicate annotation. Therefore, the tumor area was defined as the area encompassing all tumor cell nests, extending up to the most peripherally located clusters of invasive tumor cells. Subsequently, we used an automated protocol to mark the invasive margin (IM) around the tumor area (TA), with a width of 200-µm. In the individual regions of interest, we executed image analysis algorithms (IAAs) for unbiased quantification of the DAB + area and the number of stained cells in a specific ROI. In this way, we calculated the relative marker area (an estimated percentage of the area stained for a given marker in an ROI) and density (the number of stained cells per mm2 in an ROI). Depending on the type of antibody and tumor tissue, different IAAs are necessary; therefore, specific in-house IAAs for this research were developed to evaluate CD79α, CD8, FOXP3, CD163, and PD-L1 in breast cancer. The development of these IAAs is explained in more detail in the Supplementary File (Table S1, Fig. S1). In short, the IAA sampled the entire image and used threshold classification to identify DAB + pixels. To determine the threshold, different thresholds were applied to the 4 H-DAB images. One pathologist (CC) subsequently counted all positive cells for a specific staining on this image. Using this manual scoring as the gold standard, we determined which threshold yielded the least errors (i.e., missing a positive cell, counting an artifact, or counting multiple positive cells as one). Every staining used a different threshold. After staining, quality control was performed, and artifacts and areas of necrosis were excluded.

Gene expression analysis
Affymetrix gene expression data were available for a subset of patients (69 IBC and 61 nIBC) and were preprocessed as described before [25]. Briefly, raw data were normalized and batch-corrected using combat (sva package), followed by scaling of the expression data using quantile normalisation. Batch correction was required because samples originated from two different institutes (IPC, Marseille and ZAS, Antwerp). Probes with fluorescence intensities above log2(100) in at least 5% of samples were retained, after which probe redundancy (i.e., multiple probes mapping to the same gene) was resolved by selecting, for each gene, the probe with the highest standard deviation. This resulted in a final dataset comprising 19,174 unique genes. To validate immunohistochemical quantifications, correlations between mRNA expression and protein-based immune cell markers (CD8, CD163, FOXP3, CD79α, and PD-L1) were assessed using Pearson correlation coefficients. P-values inferior to 5% were considered significant. To estimate the abundance of specific immune cell populations within the tumor microenvironment, we applied consensusTME with standard parameter settings. ConsensusTME is a computational deconvolution method based on cancer-type specific consensus gene sets representing 20 distinct immune cell types [29]. Correlations between deconvolution scores and protein-based immune cell markers were calculated as described above. Furthermore, we analyzed the association between immune cell infiltration and the posterior probability of the samples in our cohort to be classified as IBC. The posterior probability scores, obtained using an elastic net classification model trained on 79 genes with IBC-specific gene expression patterns, were published previously [26]. Finally, to explore the functional role of TAMs in IBC, gene set enrichment analysis (GSEA) was performed. Patients were stratified into TAM-high and TAM-low groups using the median CD163 RMA values. Next, differential gene expression was calculated using generalized linear regression analysis (limma package) using a 2 × 2 factorial design including the tumor phenotype (i.e., IBC vs. nIBC) and the TAM stratification (i.e., TAM-high vs. TAM-low). The resulting fold-change vectors were tested for enrichment of Hallmark gene sets using normalized enrichment scores (NES–fgsea package). Both within (i.e., TAM-high vs. TAM-low in either IBC or nIBC) and between tumor phenotype (i.e., IBC vs. nIBC in either TAM-high or TAM-low) comparisons were performed to identify IBC-specific hallmarks associated with macrophage-driven tumor biology. False discovery rate (FDR)-adjusted P-values inferior to 10% were considered significant.

Statistical analysis
All statistical analyses were performed using R Studio (Version 2024.12.1 + 563). Missing data were retained in the dataset but excluded from each individual test on a case-by-case basis. Statistical significance was defined as a two-sided P-value of < 0.05. Clinicopathological differences between groups were assessed using Pearson’s chi-squared test for categorical variables and the Kruskal-Wallis test for continuous variables. To compare immune infiltration patterns between IBC and subtype-matched nIBC, logistic regression models were constructed, including immune composition variables. Correlations between immune markers (density or RMA) were analyzed using Pearson’s correlation coefficients, and differences in correlation strength were evaluated using Fisher’s r-to-z transformation. We evaluated two survival endpoints: disease free survival (DFS), defined as the interval from the date of diagnosis to the date of cancer recurrence, and overall survival (OS), defined as the interval between diagnosis and death. To dichotomize continuous immune markers, we performed receiver operating characteristic (ROC) curve analysis using the Survminer package. This method identifies the optimal cut-off point that maximizes the Youden index (sensitivity + specificity – 1), thereby improving the discriminatory power for survival outcomes. Time-to-event analyses of the dichotomized variables were performed using Kaplan-Meier estimates with comparisons via log-rank tests. The prognostic effects of immune features were first explored using univariate Cox proportional hazard models. All significant predictors from the univariate analysis were considered for multivariate Cox regression. To balance model complexity with predictive power, we used stepwise model selection based on the Akaike Information Criterion (AIC) [27]. This procedure iteratively evaluates nested models, aiming to minimize the AIC score. Lower AIC values indicate models with a better goodness-of-fit. The best-performing models were then compared to simpler base models using likelihood ratio tests (LRT) to determine if additional variables provided statistically significant improvements in model fit. Survival data were updated to December 31, 2024, and patients without events were censored at the last follow-up.

Results

Validation of IAAs
To validate our developed IAAs in Visiopharm, we analyzed the first 80 patients also with two other methods: semi-quantitative scoring by two pathologists and another software package (Definiens Tissue Studio) with pre-developed in-house algorithms. Definiens and Visiopharm image analysis software were used to quantify the RMA. In addition, in Visiopharm a density was also calculated. We performed this analysis for CD8, FOXP3, and CD163 expression. The intra-class correlation coefficient (ICC, 2-way, consistency) between the pathologists was fair (for example, CD8: 0.49, P < 0.001), but all three methods (the two software packages and the combined pathologist score) showed comparable results (Table S2). The correlations between the software packages were excellent (e.g., CD8: 0.902, P < 0.001). Furthermore, density and RMA were always strongly correlated (e.g., CD8: 0.981, P < 0.001).

Patient characteristics
A summary of the clinicopathological parameters is presented in Table 1. Besides presenting with higher stage disease (P < 0.001), inherent to the definition of IBC, no significant clinicopathological differences between the IBC and molecular subtype-matched nIBC cohort were observed. As previously demonstrated, PD-L1 expression was higher in IBC (P = 0.034), and the number of sTIL (P = 0.12) was comparable between the cohorts [8]. There were also no significant differences in the clinicopathological parameters between patients with metastatic and non-metastatic IBC (Table S3).

Composition of the immune infiltrate in IBC and nIBC
We examined the infiltration of CD8+ cytotoxic T-cells, CD79α + activated B-cells, CD163+ TAMs, and FOXP3+ Tregs in both the invasive margin (IM) and the tumor area (TA). We used the density (# immune cells/mm2) as a marker of infiltration (Fig. 2; Table 2), and validated our findings using the relative marker area (RMA, %) (Fig. S2, Table S4). We also assessed the presence of PD-L1+ immune cells. In both IBC and nIBC, TAMs were the most abundant immune cell type and Tregs were the least common. The number of immune cells in the IM was significantly higher than that in the TA for all cell types. Finally, IBC was characterized by a higher number of CD163 + TAMS in the TA (P = 0.007) and more FOXP3+ Tregs in both the IM (P = 0.04) and TA (P = 0.03) compared to nIBC.

We also compared the immune infiltrate between IBC and nIBC after dichotomization, using the median as a cut-off (Table S5). Patients with IBC had significantly more infiltration of TAMs in the TA than patients with nIBC (P = 0.015), and the number of CD8+ T-cells in the TA was higher (P = 0.044). Furthermore, the ratio of CD8+ (P = 0.050) and CD163+ (P < 0.001) cells in the TA compared to the IM, a marker for the influx of these immune cells, was significantly higher in IBC. We included all composition parameters that were significantly different between IBC and nIBC in a logistic regression model in Table 3. In this multivariate analysis, infiltration with CD163 + TAMs remained significantly higher in IBC (OR: 0.81, 95% CI: 0.70–0.94, P = 0.007), with a higher number of FOXP3+ Tregs in the IM being borderline not significant (OR: 0.85, 95% CI: 0.72–1.01, P = 0.075). This increased infiltration with CD163 + TAMs in IBC was also validated using the RMA (OR: 0.86, 95% CI: 0.75–0.97, P = 0.02).

Metastatic disease is often considered to represent late-stage disease; therefore, we also compared metastatic (IBC M1) and non-metastatic IBC (IBC M0) disease. All immune composition and clinicopathological parameters between these two groups were the same, except that patients with metastatic disease were less frequently observed to have high infiltration with sTIL (> 40%, OR: 0.78, 95% CI 0.61–0.99, P = 0.05). When comparing the non-metastatic IBC (IBC M0) with the nIBC cohort (all M0 patients), there was a significantly higher infiltration of TAMs (OR: 0.82, 95% CI 0.70–0.95, P = 0.008), which was consistent with findings in the overall IBC cohort (Table S6).

Correlations between the immune parameters
A strong immune response seems to elicit infiltration of all types of immune cells, as there is a significant correlation between all different cell types and between the number of immune cells in both the IM and TA (Fig. 3). This was confirmed using RMA (Fig. S3). Furthermore, the RMA or density of immune cells did not depend on the size of the examined area. Fisher’s r-to-z transformation was used to determine whether a specific pair of immune cell types (e.g., CD8 and FOXP3) exhibited a significantly stronger correlation than another pair involving the same immune cell type (e.g., CD8 and CD163). In IBC, the strongest correlation was observed between lymphocytes: cytotoxic T-cells and Tregs, cytotoxic T-cells and B-cells, or Tregs and B-cells (Table 4). Interestingly, there was also a strong correlation between TAMs and Tregs, but not with other lymphocytes. PD-L1 expression showed significantly stronger associations with cytotoxic T-cell (P = 0.016), Treg (P = 0.014), and TAM (P = 0.025) infiltration than with B-cell infiltration. Using RMA, we confirmed these results (Table S7). In the nIBC cohort, the strong correlation between different lymphocytes and between TAMs and Tregs was also observed, but unlike the IBC cohort (Table 4), no correlation between a pair of immune cells (e.g. CD163 and FOXP3) was significantly stronger than another pair. Finally, we assessed whether the correlation between specific immune cell pairs was significantly different in IBC compared to nIBC, but only the correlation between CD79α + B-cells and PD-L1 expression was significantly weaker in IBC (IBC 0.323 (IBC) vs. 0.612 (nIBC), P = 0.02) (Table S8).

Response to NACT and the Spatial immune composition
Response to NACT data was available for 108 IBC patients, of which 29% had a complete pathological response (pCR, n = 31/108). A higher number of sTIL (P = 0.005) and PD-L1 overexpression on immune cells (P = 0.013) were associated with pCR in IBC, as previously demonstrated [8]. A negative ER status was borderline not significant (P = 0.077). A higher number of CD8+, FOXP3+ and CD163+ cells in the TA, but not the IM was associated with pCR (Fig. 4, Fig. S4). The number of CD79α+ B-cells did not correlate with response to NACT. Next, we constructed a multivariate model (full model) including the density (or RMA) of CD8, CD163, and FOXP3 in the TA, but none of the results were significant. We also compared this full model with a base model that included only TILs, PD-L1 and ER. The AIC of the full model was higher than that of the base model (96.3 vs. 94.8), and there was no significant difference between the two models (P = 0.22), this indicated no additional benefit of adding more immune parameters besides TILs and PD-L1 to predict pCR after NACT.

Prognostic effect of the immune composition
Median follow-up in the IBC cohort was 10.46 years (95% CI: 7.43–13.52 years). The median OS in the total cohort was 5.95 years (95% CI: 4.04–14.3 years) with a 5-year OS rate of 52.6%. Patients with non-metastatic disease had a significant better prognosis with a median OS of 14.3 years (95% CI: 8.07–NR years) and a 5-year OS rate of 63.8%. Finally, median DFS was 7.14 years (95% CI: 4.07 – NR years).
The presence of CD8 + cytotoxic T-cells in the TA tended to be associated with better OS (Table 5; Fig. 5), (HR: 0.64, 95% CI: 0.4–1.03, P = 0.065). The presence of cytotoxic T-cells in the IM was associated with worse outcomes (HR: 1.75, 95% CI: 1.01–3.02, P = 0.043). Interestingly, CD79α + B-cells showed the opposite trend: their presence in the IM was associated with longer OS (HR: 0.53, 95% CI: 0.3–0.92, P = 0.022). Because of these significant prognostic differences between the location of the immune cells in the TA or IM, we also examined the TA/IM ratios (Table 5). Patients with a high TA/IM ratio of CD8 + cytotoxic T-cells had a better outcome (HR: 0.62, 95% CI: 0.39–0.99, P = 0.044), whereas the opposite was true for CD79α + B-cells (HR: 2.44, 95% CI: 1.39–4.28, P = 0.001). Although the presence of TAMs in both the TA and IM showed a trend towards shorter OS, the TA/IM ratio indicated that a greater influx of TAMs from the IM to the TA corresponded with better OS (HR: 0.56, 95% CI: 0.34–0.94, P = 0.025). Finally, since CD8 + cytotoxic T-cells are often considered to be one of the main effectors of an anti-tumoral immune response, we also examined the ratio of CD8 versus the other types of immune cells in the same ROI (Table 5). Having more CD8 + T-cells than FOXP3+ Tregs (a CD8/FOXP3 ratio above 7.8) was prognostic for longer OS (HR: 0.44, 95% CI: 0.21–0.91, P = 0.026), while solely a lower density of FOXP3+ Tregs was not. Furthermore, a high CD8/CD163 and CD8/CD79α ratio also showed a trend towards longer OS.
To further explore these findings, we included all these new significant composition variables in a multivariate model. We previously reported the association between OS and the number of sTIL (> 10%), HR status, cN-stage, and cM-stage [8]. We considered this our base model. Menopausal status, which was also prognostic in this IBC cohort (P = 0.002), was added to the model accordingly. Our complete model included all significant density variables (Density of CD8 in the IM; Density of CD79α in the IM; the TA/IM ratio of CD8, CD163 and CD79α; and the CD8/FOXP3ratio in the TA). Finally, stepwise variable selection (both forward and backward) based on the AIC, which balances the model fit with complexity, was performed to identify the optimal model. At each step, variables were added or removed, depending on whether they improved the overall model AIC. The optimal model performed significantly better than the base model (P < 0.001) and showed that besides more TILs, a positive HR status and the absence of metastatic disease; the influx of CD8 + cells (HR: 0.40, 95% CI: 0.22–0.74, P = 0.004), low infiltration of CD79α+ cells (HR: 3.59, 95% CI: 1.85–6.96, P < 0.001), and a high CD8/FOXP3 ratio in the TA (HR: 0.32, 95% CI: 0.13–0.82, P = 0.017) were all associated with a longer OS (Fig. 5).

Earlier data showed that PD-L1 was not associated with outcome in this cohort of IBC patients; however, this new composition data shows that the prognostic effect of PD-L1 might also depend on spatial expression. PD-L1 expression in the IM was associated with poor prognosis, while PD-L1 expression in the TA had a good prognosis. This was also reflected in the prognostic benefit of the high TA/IM ratio (Table S9). When these parameters were added to the best model (Table S9), the density of PD-L1 in the TA remained a significant good-prognosis factor (HR: 0.13, 95% CI: 0.03–0.56, P = 0.006), whereas the density of PD-L1 in the IM remained a poor-prognosis factor (HR: 8.12, 95% CI: 1.93–34.19, P = 0.004).
Using RMA instead of density showed very similar results (Table S10, Fig. 5), with more pronounced negative prognostic effects of TAMs (both in the TA and the IM) and CD79α + B-cells (in the TA, but not the IM). A multivariate model (Table S11) was constructed using the same stepwise approach as density. Similar to density, TILs, positive HR status, absence of metastatic disease, influx of CD8 + cells, and no infiltration of CD79α + cells were also associated with longer OS. However, patients with more TAMs (HR: 3.35, 95% CI: 1.64–6.85, P < 0.001) in the IM and more CD79α + B-cells (HR: 2.36, 95% CI: 1.16–4.79, P = 0.017) in the TA also had significantly shorter OS in the RMA analysis.
We repeated the OS analysis also for IBC patients without metastatic disease at the time of diagnosis and found very similar results. Based on our previous study [8], we included pCR, N-stage, and HR status in the multivariate model and added menopausal status (P = 0.02). We included again all composition parameters that were significant in the univariate analysis (Table S12) and followed the same stepwise approach using the best AIC to define the optimal multivariate model (Table S12). Also in the non-metastatic IBC cohort, the influx of CD8 + cytotoxic T-cells was associated with longer OS (HR: 0.32, 95% CI: 0.13–0.80, P = 0.016), while the opposite was true for the influx of CD79α + B-cells (HR: 3.73, 95% CI: 1.46–9.55, P = 0.006).
Finally, we examined DFS (Table 6, Fig. S4). In the univariate analyses, several immune cell densities and their spatial ratios were found to be associated with DFS. A high density of CD163 + TAMS in the IM was significantly associated with worse DFS (HR: 2.06, 95% CI: 1.04–4.05, P = 0.033), while higher CD79α + B-cell density in the IM was linked to improved DFS (HR: 0.37, 95% CI: 0.18–0.74, P = 0.003). Among the spatial ratios, a higher CD8/ CD79α ratio in the TA (HR: 0.57, 95% CI: 0.33–0.99, P = 0.043) and a higher CD8/FOXP3 ratio in the TA (HR: 0.44, 95% CI: 0.25–0.77; P = 0.003) were both associated with improved DFS. Similar to the findings for OS, a greater influx of CD8 + T-cells (HR: 0.45, 95% CI: 0.22–0.91, P = 0.023) and less influx of CD79α + B-cells were associated with prolonged DFS. However, the association with lower B-cell infiltration showed only borderline significance (HR: 2.92, 95% CI: 0.9–9.45, P = 0.061).
To evaluate the independent prognostic value of these immune features, a multivariate Cox proportional hazard regression model was constructed. Immune variables that showed significant univariate associations (P < 0.05) were added to the full model and stepwise variable selection was performed to identify the optimal model (Table 7). In multivariate analysis, several immune features remained significantly associated with DFS after adjusting for key clinical variables, including clinical nodal status (P = NS), menopausal status (P = NS), HR status (HR: 0.32, 95% CI: 0.15–0.71, P = 0.005), and pCR after NACT (HR: 0.30, 95% CI: 0.11–0.77, P = 0.013). A higher density of CD79α + B-cells in the IM remained a predictor of better DFS (HR: 0.14, 95% CI: 0.05–0.37, P < 0.001), while elevated TAM density in the IM was significantly associated with worse DFS (HR: 3.00, 95% CI: 1.34–6.72, P = 0.007). Furthermore, similar to OS, greater infiltration of CD8 + cells (HR: 0.45, 95% CI: 0.22–0.99, P = 0.048) was also associated with a better DFS. Similar to OS, we created a multivariate model that included spatial PD-L1 parameters. PD-L1 expression in the IM seemed to be associated with shorter DFS (HR: 3.2, 95% CI: 1.1–9.28, P = 0.024). This was also reflected in the beneficial prognostic effect of a high TA/IM ratio (HR: 0.26, 95% CI 0.08–0.86, P = 0.018). However, when added to the optimal multivariate model, these PD-L1 parameters were not significant.

Finally, we also performed DFS analysis using RMA (Table S13), with very similar results to density. In multivariate analysis, a higher RMA of CD79α + B-cells in the IM remained a predictor of better DFS (HR: 0.25, 95% CI: 0.09–0.66, P = 0.005), while more CD8 + T-cells in the IM were significantly associated with worse DFS (HR: 3.17, 95% CI: 1.38–7.30, P = 0.007). Notably, the number of TAMs in the IM group was also significantly associated with DFS. (HR: 3.10, 95% CI: 1.30–7.40, P = 0.011) and a significantly improved DFS was seen for patients with lower CD8/CD79α ratios in the TA (HR: 0.30, 95% CI: 0.11–0.58, P = 0.001).

Validation of the IHC findings with gene expression data
Affymetrix data were available for 69 IBC and 61 nIBC patients to validate our immunohistochemistry findings. First, we examined the correlations between mRNA and protein expression of all immune cell markers. We discovered significant correlations for all markers, except for FOXP3: CD8 (ρ: 0.420, P < 0.001), CD163 (ρ: 0.357, P < 0.001), CD79α (ρ: 0.561, P < 0.001) and PD-L1 (ρ: 0.285, P = 0.05). Subsequently, we used consensusTME, a deconvolution method to analyze the immune cell composition based on the expression of immune cell-related genes [28] both in the IBC and nIBC cohort (Fig. S5). For all immune cell types, we observed strong correlations between protein expression and consensusTME scores (Fig. S6 A–D). Correlations with deconvolution-based scores were also higher than those with individual gene markers, particularly for low-abundance cell types such as Tregs, supporting the use of gene sets over single marker genes. Interestingly, CD163 protein expression correlated with both M1 and M2 macrophage signatures in the IBC cohort, with a stronger association observed for M1 macrophages (Fig. S6 E).

IBC signature score and immune cell infiltration
Next, we evaluated the relationship between the IBC signature score [25] with the presence of various immune cells. In the IBC cohort, a higher IBC signature score was observed in patients with more TAMs (P < 0.001) (Fig. 6). Furthermore, there also seemed to be an association between FOXP3 Tregs and the IBC signature score (P = 0.02) (Fig. 6). Interestingly, using the consensusTME, the strongest correlation was seen between the IBC signature score and M2 macrophages; however, there was also a strong correlation with M1 macrophages and monocytes, while the correlation with Tregs was very low. In the nIBC cohort, no correlation was observed between immune cells and the IBC signature score.

The role of TAMs in the IBC phenotype
Finally, we examined the differences in gene expression and hallmark pathways for patients with a higher density of TAMs versus those with a lower density of TAMs, since the presence of TAMs seemed to be associated with an IBC phenotype. In IBC, CD163 expression was associated with allograft rejection, inflammatory response, IL-6/JAK/STAT-3 signaling, IFN-γ response, and TNF-α signaling via NF-κB. Lower CD163 expression was associated with ER response (both early and late), protein secretion, PI3K/Akt/mTOR signaling, and DNA repair (Table S14 A). In nIBC, TAMs appeared to be associated with almost the same pathways (Table S14 B). Therefore, it seems that TAMs are at least partially associated with the same inflammatory pathways in both IBC and nIBC. However, there were also significant differences between IBC and nIBC cases with higher TAM infiltration (Table 7). This table shows that E2F targets (NES = 2.156) were upregulated in IBC samples with more TAMs (compared to nIBC samples with more TAMs). This pathway contains genes regulated by E2F transcription factors that are crucial for cell cycle progression and often indicate higher proliferation. Other upregulated cell cycle-related pathways were the G2M Checkpoint (NES = 2.018) and the Mitotic Spindle pathway (NES = 1.770).

Discussion
In this study, we aimed to characterize how the anti-tumor immune response in IBC differs from that in molecular subtype-matched nIBC, and to explore its potential contribution to the aggressive clinical behavior of IBC. To do this, we used automated DIA of patient-derived pre-treatment slides stained for CD8, CD163, CD79α, FOXP3, and PD-L1. Our scoring method was validated by two pathologists and an independent software package, demonstrating that it is a reproducible and accurate tool for immune cell quantification. Depending on the staining pattern, either cell density (#/mm²) or relative marker area (RMA, %) provided a more accurate quantification of immune cells. Therefore, we chose to report both metrics, which showed a strong correlation.
We observed that TAMs were the most abundant immune cell type, Tregs were the least frequent, and immune infiltration was consistently higher in the IM than in the TA across both IBC and nIBC. However, IBC demonstrated a significant enrichment of CD163 + TAMs in the TA and FOXP3+ Tregs in both the IM and TA compared to nIBC, emphasizing the more immunosuppressive microenvironment [29]. In the multivariate model, infiltration of CD163 + TAMs remained significantly higher in IBC (OR: 0.81, 95% CI: 0.70–0.94, P = 0.007), and these differences persisted even when restricting the analysis to non-metastatic IBC (M0). Other researchers have pointed out the abundance of TAMs and their potential role in the aggressive behavior of IBC [15, 30]. To the best of our knowledge, this is the largest IHC dataset to date, which additionally makes a direct comparison with nIBC to prove the presence and importance of CD163 + TAMs in IBC. Transcriptomic deconvolution using consensusTME further reinforced these findings. Moreover, a higher IBC signature score was associated with increased TAM and Treg infiltration, a pattern absent in nIBC.
The role of Tregs in both IBC and nIBC is poorly understood. The prognostic role might be influenced by molecular subtype, spatial location, or relative abundance compared to CD8 + T-cells [31, 32]. GSEA revealed that higher infiltration of CD163 + TAMs was associated with the activation of similar inflammatory pathways in both IBC and nIBC, including TNF-α signaling via NF-κB which can promote tumor cell survival [33]. Interestingly, while overlap was observed in inflammatory pathway activation between IBC and nIBC, several inflammatory and cell-cycle related transcriptional programs were significantly more enriched in TAM-high IBC tumors than in TAM-high nIBC tumors. These findings suggest that specific IBC-associated TAMs may promote cell proliferation and invasion [29, 34]. However, it remains unclear whether IBC cells initiate TAM reprogramming or whether pre-existing macrophage polarization contributes to the establishment of the IBC phenotype. IBC tumor cells secrete a range of cytokines and chemokines (e.g., CCL2, IL-6, IL-8, …) that actively recruit and polarize macrophages towards pro-tumoral phenotypes [18, 20]. In turn, IBC-associated TAMs produce elevated levels of cytokines (e.g., CCL2, IL-6, IL-10, TNF-α, … ), which are significantly higher than those observed in non-IBC, thereby promoting tumor progression and dampening anti-tumor immune responses [17, 22, 35, 36]. Finally, macrophages in IBC can increase protease expression and facilitate the motility and invasion of established IBC cell lines [37, 38]. CD163 is considered a marker of M2 polarized TAMs, although the traditional concept of M1/M2 polarization is an oversimplification [39]. Interestingly, CD163 expression correlated with both M1- and M2-like macrophages in IBC, with a stronger association to M1-like profiles based on consensus TME analysis. This likely reflects the high plasticity of IBC-associated TAMs. In IBC, these macrophages display mixed activation features, secreting TNF-α (M1 associated) and IL-10 (M2 associated) [17], while IL-10 may also upregulate CD163 expression [35]. Moreover, IBC cells can induce macrophages with predominantly M2 characteristics yet retaining M1-associated markers and macrophages enhance endothelial sprouting and matrix remodeling in IBC regardless of their polarization state [18, 37]. Together, these findings suggest that CD163 + TAMs in IBC reflect hybrid activation states rather than a classical M2 phenotype. Future spatial transcriptomic studies could help unravel the role of these IBC-associated TAMs.
The strong correlations between various immune cell subsets, especially lymphocytes, suggest that a robust immune response drives co-infiltration. Interestingly, the strong correlation between TAMs and Tregs, that was only seen in the IBC cohort, further supports the idea of an immunosuppressive environment in IBC and potentially reflects shared immunosuppressive cytokine pathways [29]. In a preclinical breast cancer model, TAMs were shown to convert conventional CD4 + T-cells into FOXP3+ Tregs through PD-1 and TGF-β signaling [40]. In IBC, TAMs also produce CCL2 [20], a known chemoattractant for Tregs, which is likely to contribute to their co-infiltration. Moreover, EGFR activation in IBC promotes a TAM and Treg-enriched, immunosuppressive microenvironment [41].
Specific and spatially dependent immune responses were associated with survival outcomes. Increased infiltration of CD8 + T-cells in the TA was associated with improved OS, whereas the influx of CD79α + B-cells in the TA had a negative impact. Conversely, a higher number of B-cells in the IM correlated with better survival. Furthermore, the CD8/FOXP3 ratio within the same ROI proved to be a superior prognostic indicator over absolute Treg density, highlighting the importance of local immune balance. Most immune cell markers showed consistent associations with both DFS and OS, particularly CD79α in the IM and the ratio CD8 TA/IM, which were all significant across RMA and/or density analyses and retained significance in multivariate models. In addition to the infiltrating immune cells, PD-L1 expression also demonstrated a spatially dependent prognostic pattern. Although our previous study did not link PD-L1 expression on immune cells to clinical outcomes [8], our analysis revealed that expression in the TA was associated with favorable overall survival, whereas expression in the IM was linked to worse prognosis. Interestingly, while TAMs contribute to the aggressive phenotype of IBC, prognosis within IBC appears mainly driven by the spatial distribution and relative abundance of CD8 + and CD79α + cells. Although higher CD163 + TAM numbers in the IM (especially by RMA) were independently linked to poorer outcome, a higher TA/IM ratio of CD163 + TAMs correlated with better OS in univariate analysis. This could reflect a redistribution of macrophages towards the TA in the context of broader immune infiltration and enhanced anti-tumor immunity, which may also explain why this association was not retained in multivariate analysis once other immune variables were considered.
CD8 + T-cells have been associated with better outcomes in both IBC and nIBC [10, 42]. However, as was demonstrated before, their prognostic role depends on both co-infiltration with other immune cells and spatial location. CD8 + cells located near tumor cells, rather than restricted to the invasive margin, correlated with longer survival and better therapy response [15, 42–44]. While CD20 + B-cell infiltration has previously been associated with improved outcomes in IBC [13], our findings contrast this observation, as higher levels of CD79α + B-cell infiltration were linked to worse prognosis. This finding could reflect differences in the cell populations captured by each marker. CD79α stains a broader spectrum of B-lineage cells, including plasma cells, which may play distinct roles within the TME of IBC [45]. CD79α expression is also described on immature myeloid cells contributing to their tumor-promoting effects [46]. Furthermore, when the distinction between TA and IM is not made, the potentially negative impact of B cells in the TA could be masked by the beneficial prognostic effect of increased B-cell presence in the IM. B-cells organized in tertiary lymphoid structures, often located in the IM [47], are associated with a better prognosis and response to immunotherapy [48]. This can be in contrast with B-cell infiltration in the TA. In TNBC, better outcomes were associated with more numerous and spatially dispersed lymphocyte clusters containing B-cells, but a higher relative proportion of B-cells within these clusters correlated with poorer prognosis [49]. Moreover, early B- cell infiltration in TNBC can induce IL-1β secretion, which enhances invasion and migration via NF-κB signaling [50]. Finally, beyond spatial heterogeneity, interaction with TAMs and other immune cells can also shape B-cell function. Using multiplex immunofluorescence, Badr et al. demonstrated that CD20 + B-cells were associated with response to therapy, but not when CD68 + TAMs were also present [15].
Future studies should investigate the spatial relationships between TAMs, T-cells, and B-cells to identify potential therapeutic targets in IBC. Several potential targets within the immunosuppressive IBC TME have already been described [29, 34], although clinical data remain limited. To date, pembrolizumab in triple-negative disease is the only clinically approved immunotherapy for IBC, consistent with the observed PD-L1 overexpression. However, in the KEYNOTE-522 trial, only 17 patients (1.4%) with early-stage triple-negative IBC were enrolled [51] and real-world data from the INCORPORATE cohort demonstrated poor outcomes in patients with metastatic TN-IBC despite the treatment with immune checkpoint inhibitors [52]. PD-L1 expression in IBC is known to correlate with other immune checkpoint molecules such as CTLA-4, LAG3, and TIM-3 [10], although trials investigating dual checkpoint inhibition (e.g., ipilimumab plus nivolumab) have unfortunately been closed early due to insufficient accrual. This limited enrollment and unclear real-world benefit underscore the need for dedicated IBC-focused immunotherapy studies [29]. Other strategies to enhance anti-tumor immunity are also under investigation. Wang et al. demonstrated that EGFR-targeted therapy with panitumumab can shift the IBC microenvironment from immunosuppressive to immune-active, thereby enhancing the efficacy of anti-PD-L1 therapy [41]. In addition, macrophage reprogramming agents, including CSF1R inhibitors and CD40 agonists, are being evaluated in combination with immune checkpoint blockade across multiple solid tumors, including breast cancer, and could be promising in IBC given its marked TAM infiltration [53]. Targeting STAT3 signaling, a key driver of M2-like macrophage polarization, has also shown immune-modulatory effects; however, the phase II TBCRC 039 trial with ruxolitinib demonstrated suppression of pSTAT3 but no clinical benefit in IBC [54].
This study represents the largest IHC-based spatial immune profiling of IBC to date. However, some limitations of this study should be acknowledged. The study was descriptive and observational in nature, limiting causal inferences regarding the functional roles of specific immune subsets. Another limitation concerns the use of dichotomized immune variables. Since there are currently no validated or universally accepted cut-off points for immune cell densities or ratios in IBC, the thresholds used, whether based on cohort medians or optimal cut-offs from ROC analysis, remain somewhat arbitrary. Therefore, external validation in independent cohorts is necessary. In addition, while the IHC approach remains clinically feasible and widely used, more advanced technologies, such as multiplexed spatial transcriptomics or single-cell RNA sequencing, could provide deeper mechanistic insight into immune cell phenotypes and interactions. Nonetheless, we believe that our findings are highly relevant for future translational and therapeutic studies in IBC.

Conclusion
We explored the immune microenvironment of IBC to better understand its aggressive clinical behavior. Our findings highlight crucial differences in immune composition and spatial distribution compared to subtype-matched nIBC, with notably more infiltration of CD163 + TAMs in IBC. Gene expression analyses validated our immunohistochemical findings, and pathway analysis linked high TAM density to inflammation, immune suppression, and cell proliferation pathways. Together, this illuminates the role TAMs could play in fostering the aggressive IBC phenotype, and targeting macrophage polarization or cytokine-mediated pathways could be new therapeutic options specific to IBC. Furthermore, we demonstrated the clinical importance of immune cell localization beyond absolute quantities. Spatial context determines the clinical implications of CD79α + B-cell and CD8 + cytotoxic T-cell infiltration in IBC. A greater influx of cytotoxic T-cells into the TA positively correlated with improved OS, while the influx of B-cells was associated with a worse prognosis. This underscores the necessity of considering spatial context when evaluating immune cell biomarkers, and therapeutic strategies aimed at modulating immune cells should carefully assess spatial dynamics.

Supplementary Information
Below is the link to the electronic supplementary material.