Spatial transcriptomics reveals tumor microenvironment–driven subtypes of invasive lobular carcinoma

Significance

The tumor microenvironment (TME) of invasive lobular carcinoma (ILC) remains largely unexplored despite its clinical relevance and distinct biology. Using spatial transcriptomics, we reveal insights into ILC spatial architecture and identify four tumor microenvironment subtypes characterized by specific biological features, including androgen receptor signaling, metabolic pathways, and M2-like macrophage infiltration. This classification provides refined prognostic stratification and suggests broad therapeutic avenues. Our approach integrates morphological annotation, spatial gene expression, and single-cell deconvolution to resolve ILC heterogeneity at multiple levels. We further validate our findings in large-scale bulk transcriptomic cohorts. This multimodal framework offers a comprehensive view of ILC biology and establishes a foundation for microenvironment-informed precision oncology in this neglected breast cancer subtype.

Abstract

Invasive lobular carcinoma (ILC) is the second most common histological subtype of breast cancer and displays distinct clinical and biological behavior compared to breast cancer of no special type. However, current molecular classifications largely overlook its complex spatial organization and tumor microenvironment (TME). Here, we performed spatial transcriptomics on 43 hormone receptor-positive, HER2-negative (HR+/HER2−) ILC tumors with detailed morphological annotation and long-term clinical follow-up. By integrating spatial gene expression with histology and single-cell deconvolution, we characterized the composition and architecture of the TME and revealed high inter- and intratumor heterogeneity. Spatial clustering uncovered cell populations and pathways linked to clinical outcome. We then developed a multimodal classification of ILC by integrating gene expression, morphology, and spatial metrics, identifying four distinct subtypes: normal/stroma-enriched (NSE), proliferative (P), androgen receptor-enriched (ARE), and metabolic/immune-enriched (MIE). These subtypes, collectively termed ILC4TME, reflect the interplay between tumor and microenvironmental features. Gene signatures derived from the spatial data enabled subtype assignment in external bulk RNA-seq and microarray datasets (SCAN-B, METABRIC), revealing reproducible biology and significant associations with survival. In multivariable models, ILC4TME retained prognostic value beyond established gene signatures and clinicopathological variables. Notably, the P subtype was linked to poor prognosis, even in patients treated with endocrine therapy alone, while the NSE subtype was associated with favorable outcomes. Our findings uncover spatial and cellular heterogeneity in ILC that is not captured by existing classification approaches, offering a refined framework for risk stratification and therapeutic targeting based on tumor microenvironment architecture.

Invasive lobular carcinoma (ILC) accounts for 10 to 15% of all breast cancers (BCs) and is the most common special type of BC. Mainly positive for hormone receptors and negative for ERBB2 amplification (HR+/HER2−), most of these tumors usually belong to the luminal intrinsic molecular subtypes of BC (1). Single-file pattern of cancer cells and the lack of E-cadherin adhesion molecule are key features of this histological subtype. The loss of E-cadherin can result from the inactivation of the CDH1 gene by mutations, promoter methylation, or copy number loss of the 16q22 chromosomal region including CDH1.

ILC shows distinct metastatic patterns compared to breast cancer of no special type (NST), with spread to peritoneum, retroperitoneum, gastrointestinal and gynecologic organs, and leptomeninges, in addition to bone. Standard treatment includes surgery, endocrine therapy, radiotherapy, and chemotherapy (2–6), but the low proliferation of ILC limits chemotherapy benefit (7). Prognosis is associated with age, nodal status, tumor size, Ki-67, and hormone status (8). Importantly, ILCs are more likely to present late relapse than NST BCs, and previous studies demonstrated that the prognosis of ILC patients is worse than that of NST patients (9). Consequently, the complexities and unique challenges associated with ILC necessitate a concerted effort toward more extensive and focused research to better understand its biology, improve early detection methods, and develop targeted therapies.

Widely used prognostic signatures [such as Oncotype DX (10), MammaPrint (11), EndoPredict (12)] were not developed for ILC specifically. Next to the commercial tests, the Genomic Grade Index (13) (GGI), which captures tumor cell proliferation, showed a prognostic value in ILC as well (14). LobSig (15) is the only gene signature that was developed for patients with ILC, but additional validation is needed. Proliferation and immune-related gene signatures from gene expression analyses define 3 ILC subtypes, namely proliferative, immune-related, and reactive-like (16). However, this classification was derived using bulk RNA sequencing data without considering the proportion and the spatial organization of tumor microenvironment (TME) cell types.

Regarding the TME, ILC generally has low lymphocytic infiltration (median TILs ~5%), yet higher TILs associate with younger age, nodal involvement, proliferation, and poorer prognosis (17). Moreover, macrophages were found to play a role in shaping immune response in HR+ BC, showing significant differences in populations between NST BC and ILC (18).

Spatial transcriptomics (ST) has emerged as a novel technique that allows RNA sequencing with spatial resolution in individual tissue sections, addressing the challenge of characterizing intratumor heterogeneity and cell–cell interactions, well-known limitations of bulk RNA sequencing (19). Moreover, combining spatial transcriptomics with single-cell RNA sequencing data significantly improves the resolution of gene expression mapping.

With the aim of investigating ILC heterogeneity including its microenvironment architecture, we applied spatial transcriptomics (ST) on a well-annotated series of 43 HR+/HER2− ILC samples. In particular, we analyzed gene expression and imaging data from ST, which we then integrated to develop a four-subtype classification of ILC capturing tumor intrinsic and TME features, hereafter referred to as ILC4TME. Recognizing that ST is not applicable for routine clinical use, we further developed a gene expression-based method for the assignment of our ST-derived ILC subtypes from bulk RNA sequencing data, which we validated in the METABRIC (20–22) and SCAN-B (23) cohorts.

By combining spatial transcriptomics to define subtypes within the tissue context, single-cell data to resolve their cellular composition, and bulk RNA sequencing to validate findings across large clinical cohorts, we leveraged the complementary strengths of each modality to achieve a comprehensive and translatable characterization of ILC heterogeneity.

Results

Spatial Gene Expression Profiling of Primary ILC Samples.

To decipher the composition and organization of ILC and its microenvironment, we performed spatial transcriptomics (ST, Visium–10× Genomics) on 43 HR+/HER2− primary ILC frozen surgical samples from patients with well-annotated clinicopathological data and long-term follow-up (median follow-up of 12.1 y for relapse-free interval; see Dataset S1). No neoadjuvant therapy was administered. Among these patients, nine experienced disease relapse (median relapse time of 5.81 y). At diagnosis, most of the tumors (35 tumors) were classified as classic, four as pleomorphic, three as mixed, and one as the solid histopathological subtype. The complete workflow of the study is shown in Fig. 1. In total, 71,997 ST spots were analyzed, with a median of 1,549 spots per array and a median of 1,300 genes at the spot level.

Fig. 1.

Overview of the study. Spatial transcriptomics (ST, Visium-10× Genomics) was performed on 43 primary ILC tumors collected at surgery. Hematoxylin and eosin (H&E) slides relative to ST samples were annotated by a trained pathologist assisted by an AI-based tool, and co-occurrence analysis between morphological classes was conducted. ST spots were clustered based on their gene expression similarity and further deconvoluted using external single-cell RNA sequencing data. Morphology and gene expression information were then integrated with the aim of obtaining a TME-based classification of ILC, that was validated in external bulk RNA sequencing datasets.

Morphological and Transcriptomic Heterogeneity Reveals Distinct Tumor Microenvironment Interactions in ILC.

Invasive lobular carcinoma (ILC) exhibits pronounced morphological and molecular heterogeneity, reflected in both the composition of its microenvironment and the spatial organization of tumor and stromal components. High-resolution H&E slides corresponding to each spatial transcriptomics (ST) sample were annotated into 12 morphological classes by a dedicated breast pathologist using a machine learning-based digital pathology approach in QuPath (24) (Fig. 2 A and B). This framework enabled systematic identification of tumor, stromal, and immune cells, as well as additional histological features including high lymphocyte density stroma (semiquantitatively defined regions with ≥50% lymphocyte density), cellular (fibroblasts)/acellular (extracellular matrix) stroma, adipose tissue, in situ carcinoma, normal breast tissue, vessels, nerve, apocrine metaplasia and microcalcification (see SI Appendix, Materials and Methods for details).

Fig. 2.

Morphological and transcriptomic heterogeneity of ILC. (A) H&E slide from sample ST64. (B) Annotated H&E slide relative to sample ST64 with labels showing the different morphological categories. (C) Spatial projection of ST spot clusters for sample ST64. (D) Enrichment (green) and depletion (orange) of major tier cell types from single-cell deconvolution in different morphological compartments (adjusted P < 0.25). (E) Heatmap showing gene set variation analysis (GSVA) for Hallmark of Cancer across the ST spot clusters. Clusters were labeled as Tumor/In situ or TME if they were enriched (from morphology) in tumor cells/in situ carcinoma or tumor microenvironment cell types, respectively. (F) Spatial representation of stroma-enriched spots (from annotation, Left) and relative abundance of CAFs cell type (from CARD single-cell deconvolution, Right) at the spot level in sample ST64. (G) Enrichment (green) and depletion (orange) of minor tier cell types from single-cell deconvolution in the different ST spots clusters (adjusted P < 0.25). (H) Boxplots showing the association between clusters percentages and relapse status (*: adjusted P < 0.05, **: adjusted P < 0.01, ***: adjusted P < 0.001). Only clusters shared across a substantial number of samples are shown. P-values were from the two-sided Wilcoxon rank-sum test and corrected using the Benjamini–Hochberg procedure.

Analysis of the relative abundance of these morphological classes revealed substantial interpatient variability (SI Appendix, Figs. S1 and S2). Notably, younger patients (<50 y) showed a significant enrichment of high-TIL stroma and normal breast tissue (adjusted P = 0.0034 and 0.0037, respectively). To better capture spatial relationships between tumor and its surrounding microenvironment, we quantified co-occurrence scores measuring the conditional probability of tumor spots overlapping with specific tissue features. This analysis revealed a significantly higher co-occurrence between tumor cells and both high-TIL stroma and normal breast tissue in younger patients (adjusted P = 0.00254 for both). These findings highlight how patient-specific factors shape the spatial architecture of ILC and its immune contexture (SI Appendix, Fig. S3 and Datasets S2 and S3).

To further dissect the microenvironmental composition at higher resolution, we deconvolved the ST spots using a single-cell RNA sequencing reference dataset, inferring the spatial distribution of tumor, immune, and stromal populations across tissue compartments using CARD (25) deconvolution software. The resulting cell-type maps showed high concordance with morphological annotations, supporting the biological accuracy of the inferred compositions (Fig. 2 D and F and SI Appendix, Fig. S4). In addition to this strong agreement, we validated the robustness of our deconvolution by comparing CARD-derived cell-type scores with results from the independent xCell algorithm applied to pseudobulk transcriptomes. These two approaches showed high concordance across major cell types (e.g., Spearman’s ρ = 0.65 for endothelial cells, ρ = 0.65 for fibroblasts, and ρ = 0.79 for macrophages), indicating that our spatial deconvolution reliably captures tumor microenvironment composition even in the absence of ILC-specific data in the reference dataset from Wu SZ et al. (26). Distinct compartments exhibited characteristic immune signatures: Adaptive immune cells (B and T cells) were enriched in normal epithelial regions, while adipose tissue was enriched in myeloid populations (monocytes, macrophages, cycling myeloid cells) alongside endothelial and perivascular-like cells. These patterns suggest compartment-specific immune organization within the ILC microenvironment.

At the transcriptomic level, unsupervised clustering of all ST spots identified 23 distinct clusters (Fig. 2 C, E, and G), with a median of 14 clusters per patient (range: 9 to 19) (SI Appendix, Fig. S5). These clusters were characterized by integrating morphology, cell-type composition, differential gene expression, and pathway enrichment analyses, revealing strong concordance across modalities (Dataset S4). Clusters enriched in adipose tissue, blood vessels, or normal glands tended to be widely shared across patients, whereas tumor-, in situ carcinoma-, and fibroblasts-enriched clusters were often more sample-specific, underscoring the diversity of tumor and stromal states within ILC.

Several transcriptionally defined clusters showed significant associations with clinical outcome (Fig. 2H and SI Appendix, Fig. S7). Clusters enriched in fibroblasts and stromal signaling (cluster 1), mixed epithelial states with high estrogen signaling and low proliferation (cluster 9), or normal epithelial and perivascular-like cells with WNT, Notch, and hormonal pathway activation (cluster 10) were associated with nonrelapse. Similarly, cluster 17, enriched in inflammatory and myofibroblastic CAFs, showed immune-related pathway activation and was linked to favorable outcomes. In contrast, cluster 15, characterized by high MYC expression and proliferative signatures, was significantly associated with disease relapse despite its relatively low frequency. Although some of these clusters, such as cluster 15, were present in a smaller subset of samples, statistical associations were evaluated using Wilcoxon rank-sum tests, which are robust to outliers and group size imbalances. Furthermore, all clusters, including those less frequent, were systematically assessed for quality metrics including UMI counts, number of expressed genes, and spatial coherence. These evaluations confirmed that even rare clusters exhibited clear spatial boundaries, reproducible transcriptomic programs, and biological interpretability (SI Appendix, Fig. S6).

Together, these analyses reveal a highly heterogeneous morphological and molecular landscape in ILC, shaped by patient characteristics and spatially organized tumor microenvironment interactions. This integrated framework highlights clinically relevant patterns of tissue composition, immune contexture, and transcriptional states, providing insights into the biological diversity of ILC and its potential impact on disease progression.

ILC Classification Based on TME Heterogeneity.

With the aim of stratifying ILC patients based on tumor architecture and their microenvironment, we integrated RNA sequencing data, spatial and morphological information obtained from ST experiments using a multimodal approach based on integrative non-negative matrix factorization (integrative NMF). The integrative NMF framework performs consensus clustering on the joint latent representation across modalities, resulting in a single discrete subtype assignment for each tumor. Three matrices—i) the fraction of each gene expression-based cluster, ii) the fraction of each annotated cell type computed as the average composition of each spot across the tissue, and iii) the level of co-occurrence between tumor microenvironment annotated cell types and tumor cells—were used as input for patient-level clustering using the intNMF (27) algorithm, implemented in the MOVICS (28) R package (Datasets S5–S7). For this analysis, morphological classes “High-TIL stroma” and “Lymphocyte” were merged into a single “Immune” class. Two samples lacking annotated tumor regions were excluded from the analysis, as the co-occurrence component of the model requires the presence of both tumor and microenvironment compartments. The Cluster Prediction Index was used to determine the optimal number of clusters, resulting in the identification of four ILC subtypes, namely normal/stroma-enriched (NSE), proliferative (P), androgen receptor-enriched (ARE) and metabolic/immune-enriched (MIE) (Fig. 3 A–C and SI Appendix, Fig. S8).

Fig. 3.

ILC4TME classification. (A) Heatmap showing the association between ILC4TME subtypes, clinical features, enrichment in ST spot clusters, morphology, and co-occurrence level between tumor cells and TME-related cell types (from morphology). (B) Heatmap highlighting normalized enrichment score (NES) of the top five enriched pathways (adjusted P < 0.25) on differentially expressed genes of the ILC4TME subtypes from DGEA on sample pseudobulks. (C) Enrichment (green) and depletion (orange) of minor tier cell types from single-cell deconvolution in the different ILC4TME subtypes (average of all the ST spots by sample, adjusted P < 0.25). (D) Kaplan–Meier plot showing RFI (in years) for ILC4TME subtypes in our ST cohort. P-value was from the log-rank test.

The NSE subtype (n = 10) was enriched in cellular stroma (from morphological annotation) and showed a higher level of co-occurrence between tumor cells, cellular/acellular stroma, and in situ carcinoma. Gene expression analysis revealed enrichment in myCAFs, and a higher presence of clusters 1, 6, 14, and 22 (stroma and inflammation-related clusters). The P subtype (n = 12) was enriched in tumor cells (according to both morphological annotation and single-cell deconvolution). It also exhibited a higher representation of clusters 4, 5, 8, 11, 13, and 19 (which are mainly associated with tumor, cycling cell types, and adaptive immunity-related clusters). The ARE subtype (n = 9) was characterized by high expression of androgen receptor and enrichment in endothelial and plasmablast cells (from single-cell deconvolution). This subtype also showed a higher presence of clusters 0 and 21 (related to stroma and epithelial/metabolic functions, respectively). Finally, the MIE subtype (n = 10) was associated with a higher presence of adipose tissue, vessels, and a greater level of co-occurrence between adipocytes and cancer cells (from morphology). Single-cell deconvolution revealed an enrichment of endothelial-PVL cells, macrophages, and monocytes. Clusters 3 and 12 were more prevalent in MIE, corresponding to stroma/adipogenesis/innate immunity and epithelial/metabolic-related clusters, respectively.

To further characterize the subtypes, we performed differential gene expression analysis using pseudobulk samples. For each subtype, we identified a substantial number of differentially expressed genes, with 105 in ARE, 605 in P, 133 in MIE, and 65 in NSE, highlighting transcriptional differences among subtypes (adjusted P < 0.05). GSEA conducted on the results of the differential gene expression analysis (Dataset S14) provided consistent results with the enrichment analysis described above (Fig. 4B). Indeed, the NSE subtype showed enrichment in epithelial–mesenchymal transition (EMT), myogenesis, and, interestingly, apical junction. The P subtype was enriched in cell cycle-related pathways. The ARE subtype was enriched in androgen response and metabolic pathways, while the MIE subtype was enriched in metabolic, adipogenesis, and immune-related pathways.

Fig. 4.

Identification of ILC4TME subtypes in the SCAN-B HR+/HER2− ILC cohort and their prognostic significance. (A) Enrichment (green) and depletion (orange) of xCell deconvoluted cell types (Left) and Hallmark of Cancer gene signatures (computed using GSVA, Right) in the different ILC4TME subtypes (adjusted P < 0.25). (B) Kaplan–Meier plot showing RFI (in years) for ILC4TME subtypes. P-value was from the log-rank test. (C) Kaplan–Meier plot showing OS (in years) for ILC4TME subtypes. P-value was from the log-rank test. (D) Forest plot summarizing univariable Cox regression analyses of RFI in the SCAN-B cohort, including ILC4TME subtypes, established gene prognostic signatures, Ciriello et al. derived subtype signatures, and clinical variables. (E) Kaplan–Meier plot showing OS (in years) for ILC4TME subtypes in the subset of patients who received adjuvant endocrinotherapy only (no chemotherapy). (F) Forest plot of different ILC4TME subtypes for RFI by multivariable analysis (controlling for prognostic signatures shown in D) in SCAN-B. (G) Forest plot of different ILC4TME subtypes for RFI by multivariable analysis (controlling for clinical features shown in (D) in SCAN-B.

In our dataset, no significant association was found between our ILC classification and clinical features such as tumor size, nodal status, and age at diagnosis, but a nonsignificant trend toward an association between P subtype, grade 3 tumors, and higher Ki-67 was observed (Dataset S15). Survival analysis revealed that the NSE subtype showed a trend toward better disease outcome (Fig. 3D).

In summary, the integration of spatial RNA sequencing, morphological data, and spatial information led to the development of the ILC4TME classification that highlights the complex interplay between tumor cells and their surrounding environment, opening broad avenues for ILC treatment tailoring.

Identification and Prognostic Characterization of ILC4TME Subtypes in Bulk RNA-seq Datasets.

To enable the identification of these subtypes from bulk RNA-seq datasets, we developed a classifier based on gene signatures. In particular, four gene signatures (one for each subtype) were derived from the pseudobulk data using differential gene expression analysis (DGEA, list of differentially expressed genes provided in the Zenodo repository, see Data, Materials, and Software Availability). For each subtype, differential expression was performed in a one-vs-rest fashion, and signatures were restricted to (maximum n = 100) positively and uniquely up-regulated genes to capture subtype-defining transcriptional programs. The final signatures included 50 genes for NSE, 100 genes for P, 82 genes for ARE, and 100 genes for MIE. To evaluate the ability to assign these subtypes to individual samples based on gene expression, we applied the four signatures to the pseudobulk of each sample. Signature scores were computed as the average expression of their constituent genes, rescaled across samples, and samples were assigned to the subtype corresponding to the highest standardized signature score. This method resulted in the misclassification of only four samples (two P samples and two ARE samples classified as NSE), achieving an accuracy of 0.90 and a Matthews correlation coefficient (MCC) of 0.88 (Dataset S8).

We next aimed to identify the ILC subtypes and validate their underlying biological characteristics in the HR+/HER2− ILC cohorts from the METABRIC (microarray data, n = 122) and SCAN-B (bulk RNA sequencing data, n = 853) datasets (Datasets S9 and S10, respectively). While METABRIC offers longer-term follow-up, SCAN-B represents a more suitable primary validation cohort for a tumor microenvironment–derived classifier due to its larger sample size and RNA sequencing profiling without tumor cellularity filtering. Nevertheless, METABRIC provides complementary validation through its extended follow-up. In the SCAN-B cohort, 291 samples were classified as NSE, 226 as P, 186 as ARE, and 150 as MIE, each subtype sharing similar characteristics to those derived from the ST analysis according to gene set variation analysis (GSVA). In particular, the ARE subtype was enriched in androgen response and metabolic-related pathways, the MIE was enriched in adipogenesis, metabolism, and immune-related pathways, the P samples showed enrichment in cell cycle and proliferation-related pathways, and the NSE subtype was enriched in extracellular matrix, stroma, and inflammation/immune-related pathways (adjusted P < 0.25, Fig. 5A). Of interest, the MIE subtype showed higher levels of adipocytes, M2 macrophages, and endothelial cells (adjusted P < 0.25), while the P subtype was enriched in CD4 and helper T cells (adjusted P < 0.25, Fig. 4A) from xCell deconvolution. Similar results were found in the METABRIC cohort (SI Appendix, Fig. S11).

Fig. 5.

Validation of ILC4TME subtypes and their concordance with histology and established ILC taxonomies. (A) Boxplot showing higher expression of the AR gene in the ARE subtype in ST cohort pseudobulk. (B) Boxplot showing higher expression of the AR gene in the ARE subtype in SCAN-B. (C) Stacked bar charts showing the percentage distribution of ILC4TME subtypes within ILC and NST histologies (HR+/HER2− BC) in SCAN-B. (D) Alluvial plots showing the level of concordance between ILC4TME classification and Ciriello et al. bulk classification in METABRIC. (E) Heatmap showing standardized residuals for AR expression from IHC (in quintiles) across ILC4TME subtypes (P = 0.0413 from Fisher’s exact test) in the ST cohort. (F) Forest plot of different ILC4TME subtypes for RFI by multivariable analysis (controlling for Ciriello et al. classification-related signatures) in SCAN-B.

We next examined associations between the ILC4TME subtypes and clinical features (Dataset S12 for SCAN-B, Dataset S11 for METABRIC). In both datasets, higher tumor grade was significantly associated with the P subtype (P < 0.001 for SCAN-B, P = 0.0187 for METABRIC). In SCAN-B, nodal involvement and larger tumor size were also significantly linked with the P subtype (P = 0.0203 and P < 0.001, respectively). Notably, a higher number of CNVs was associated with the P subtype in the METABRIC dataset (SI Appendix, Fig. S11).

Survival analyses revealed significant differences among ILC4TME subtypes in SCAN-B for both relapse-free interval (RFI, P = 0.002, Fig. 4B) and overall survival (OS, P = 0.001, Fig. 4C), with the NSE subtype associated with favorable outcome and the P subtype with poorer prognosis in univariable analysis. These associations for NSE and P remained significant after adjustment for tumor grade, tumor size, nodal involvement, and age at diagnosis (Fig. 4G and SI Appendix, Fig. S9). In METABRIC, similar trends were observed for both RFS (P = 0.062) and OS (P = 0.056), consistent with SCAN-B but not reaching statistical significance, likely reflecting the smaller cohort size (SI Appendix, Fig. S10).

To evaluate the robustness of ILC4TME across treatment contexts, we performed therapy-stratified survival analyses in SCAN-B. In patients receiving endocrine therapy alone, the P subtype was associated with significantly worse OS (HR = 2.2, P = 4.1 × 10⁻⁴; adjusted P = 0.0016), an effect that persisted after multivariable adjustment (HR = 1.7, P = 0.029), whereas the NSE subtype showed favorable prognosis (HR = 0.57, P = 0.015; adjusted P = 0.029). Similar, though weaker, trends were observed for RFI. In chemotherapy-treated patients, NSE tumors again showed favorable outcome, while the ARE subtype tended toward poorer prognosis, although these associations did not remain significant after clinical adjustment (SI Appendix, Figs. S14 and S15).

To assess the added prognostic value of ILC4TME over established mRNA-based signatures, we performed multivariable Cox regression analyses in both the SCAN-B and METABRIC cohorts adjusting for ROR, Oncotype DX, Genomic Grade Index, EndoPredict, MammaPrint, and LobSig (Fig. 4F and SI Appendix, Fig. S9). Fig. 4D compares the prognostic power (univariable analysis) of ILC4TME subtypes, clinical features, and prognostic signatures in ILC. In SCAN-B, inclusion of ILC4TME significantly improved model fit for relapse-free interval when adjusting for genomic signatures (likelihood ratio test P = 0.044; ΔC-index = +0.023) and showed a comparable improvement when adjusting for clinical variables (ΔC-index = +0.021). Improvements in overall survival models were more modest but consistent (ΔC-index = +0.014). In METABRIC, although statistical power was limited by cohort size, inclusion of ILC4TME consistently improved risk discrimination across models. When adjusting for genomic signatures, ILC4TME increased the C-index for RFS (ΔC-index = +0.007) and OS (ΔC-index = +0.009). Larger improvements were observed when adjusting for clinical variables (RFS ΔC-index = +0.035; OS ΔC-index = +0.006). Although these effects did not always reach statistical significance, their direction and magnitude were consistent with those observed in SCAN-B, supporting the reproducibility of the added prognostic information captured by ILC4TME. After adjustment for proliferation-based scores, the prognostic effect of the P subtype was attenuated, indicating that it is largely driven by proliferative activity. In contrast, the NSE subtype retained independent prognostic value, remaining significantly associated with improved survival in both datasets, demonstrating that its favorable prognosis cannot be explained solely by low proliferation. Results are shown in Fig. 4 and SI Appendix, Figs. S9 and S10 and Table S2.

To elucidate the biological basis underlying this favorable prognosis in NSE, we examined its spatial and transcriptional features. NSE tumors were characterized by a higher co-occurrence of tumor cells with stromal elements in the spatial dataset, particularly myofibroblastic CAFs, and by enrichment of apical junction pathways also observed in external cohorts, features consistent with a more organized epithelial–stromal interface and potential physical containment of tumor cells. Notably, EMT-related pathway enrichment in NSE originated predominantly from stromal rather than tumor compartments, indicating fibroblast activation and tissue remodeling rather than increased tumor invasiveness (SI Appendix, Fig. S17).

ILC4TME Signatures Highlight Lobular Carcinoma-Enriched Tumor Microenvironment Programs.

To determine whether the ILC4TME classification captures lobular-specific biology rather than general ER+/HER2− breast cancer features, we applied it to HR+/HER2− NST tumors from the SCAN-B cohort, enabling direct comparison across histologies.

We first assessed subtype prevalence using a chi-square test after discretely assigning ILC4TME subtypes to HR+/HER2− NST tumors by rescaling signature values based on ILC-derived parameters. This revealed a strong association between histology and subtype distribution (χ² = 257.77, P < 0.001). Specifically, the MIE and NSE subtypes were significantly overrepresented in ILC, whereas the proliferative subtype was enriched in NST tumors; ARE showed no strong histology bias (Fig. 5C). These results were corroborated by analyses of continuous signature expression levels and receiver operating characteristic curves, which consistently showed higher MIE and NSE signal in ILC and higher proliferative signal in NST, while the ARE program showed comparable behavior across histologies (SI Appendix, Figs. S16 and S19 and Table S1). To probe whether ILC4TME signatures extend beyond the HR+/HER2− setting, we examined HER2+ and TNBC ILC cases from SCAN-B. Despite limited sample size, the proliferative subtype was more frequent in HER2+ ILC, while the ARE subtype showed a trend toward enrichment in TNBC ILC, consistent with, respectively, the typical higher proliferation of HER2+ tumors and the luminal androgen receptor (LAR) (29) TNBC subtype. These effects were smaller than those observed between ILC and NST, suggesting that ILC4TME captures programs intrinsic to lobular carcinoma (SI Appendix, Fig. S16 and Table S1).

To assess whether ILC4TME subtypes exhibit comparable transcriptional coherence across histologies, we quantified subtype compactness using two complementary metrics. Pearson correlation between each tumor and its assigned subtype centroid was significantly higher (P < 0.001) in ILC than in NST for three of four subtypes (NSE, P, ARE), indicating more compact and well-defined programs in lobular carcinoma. The MIE subtype showed similar coherence across histologies, despite being more prevalent in ILC, suggesting preserved biological structure when present in NST (SI Appendix, Fig. S20). Consistently, the explained variance of subtype marker genes was higher in ILC than in NST, supporting greater biological resolution of ILC4TME programs in lobular tumors (SI Appendix, Table S1). To explore extension beyond the HR+/HER2− setting, we applied the same framework to HER2+ and TNBC ILC cases. Subtype coherence was largely preserved across clinical subgroups, although differences were qualitative and limited by sample size (SI Appendix, Fig. S20 and Table S1).

We next evaluated whether the prognostic impact of ILC4TME subtypes differs between histologies. A global Cox model including subtype, histology, and their interaction showed no significant interaction term, indicating no strong global modification of subtype effects by histology. However, stratified analyses revealed clear biological differences. In HR+/HER2− ILC, Kaplan–Meier analyses showed significant differences among the four subtypes for relapse-free interval (P = 0.002), with the proliferative subtype associated with the worst outcome and the NSE subtype with the most favorable outcome. Notably, within ILC, NSE and MIE displayed significant separation in RFI (Cox HR = 2.14, P = 0.041), indicating clinically relevant heterogeneity among nonproliferative, microenvironment-associated subtypes. In contrast, in HR+/HER2− NST tumors, prognostic separation was largely driven by the proliferative subtype, while NSE, MIE, and ARE showed overlapping survival curves with no evidence of separation (e.g., NSE vs. MIE: Cox HR = 0.85, P = 0.71). Together, these findings indicate that while proliferation dominates prognosis across histologies, additional microenvironment-related programs captured by ILC4TME contribute to clinically meaningful heterogeneity specifically in invasive lobular carcinoma (SI Appendix, Figs. S21 and S22). As an exploratory analysis, we investigated the prognostic power of ILC4TME signatures in other BC subtypes. However, their application to HER2+, TNBC, and HR+/HER2− NST tumors revealed only limited prognostic associations, except in the HR+/HER2− NST cohort, where certain signatures maintained prognostic value (SI Appendix, Fig. S13).

Collectively, these results indicate that ILC4TME reflects lobular-specific tumor microenvironment programs with higher transcriptional coherence and prognostic relevance in ILC than in NST tumors. This highlights the added biological and clinical value of spatially informed stratification in lobular breast cancer.

ILC4TME Reveals Unique Dimensions Beyond Established Classifications.

To contextualize the ILC4TME classification, we compared it with established histological, intrinsic, and molecular classifications. Histologically, our cohort included four pleomorphic ILC (PILC) cases, which were assigned to three different ILC4TME subtypes, two to proliferative, one to normal/stroma-enriched, and one to metabolic/immune-enriched, highlighting the heterogeneity of these variants. Mixed ILC cases and the single solid ILC case were also distributed across multiple subtypes. This diversity underscores the added value of transcriptome- and TME-based stratification over histological definitions.

In terms of intrinsic subtypes, P tumors were enriched in luminal B and HER2-enriched profiles, while NSE tumors aligned more often with the normal-like group (P < 0.001 in both SCAN-B and METABRIC, Datasets S11 and S12). We also assessed overlap with the Ciriello bulk-derived classification. In METABRIC, the association between ILC4TME and the Ciriello subtypes showed a nonsignificant trend (P = 0.082, Fig. 5D and Dataset S11), but residuals indicated alignment between P (from ILC4TME) and the proliferative group (from Ciriello et al.) and between MIE and the immune-related group. To compare the two classifications, we extended the ILC4TME classification to the subset of HR+/HER2− ILC samples in TCGA (30) (n = 99). However, TCGA was used exclusively for this comparison due to limitations in the quality of its clinical and survival data. In this cohort, the association was statistically significant (P < 0.01; SI Appendix, Fig. S12 and Dataset S13). In our ST dataset and in the SCAN-B cohort, where Ciriello subtypes are not directly available, we computed the expression levels of their immune-related and reactive-like gene signatures across the ILC4TME subtypes. As expected, the MIE subtype showed the highest expression of the immune-related signature, while NSE and ARE were enriched for reactive-like features (Fig. 5 A and B and SI Appendix, Fig. S18). Since the proliferative group in the Ciriello classification was defined by low expression of both immune-related and reactive-like programs, it was not linked to a distinct gene set and thus could not be evaluated in this context. Importantly, the prognostic value of ILC4TME remained significant after adjustment for the immune-related and reactive-like signatures defined by Ciriello et al. in SCAN-B (likelihood ratio test P = 4.1 × 10⁻⁴ and P = 0.0086), with consistent but nonsignificant trends observed in METABRIC (Fig. 5F and SI Appendix, Fig. S10).

Our classification also introduces the ARE subtype, absent in prior frameworks, which shows higher AR gene expression in both our dataset and SCAN-B (P = 0.011, P < 0.001, respectively, Fig. 5 A and B) and enrichment in androgen receptor signaling. Likewise, MIE refines the immune-related group by capturing spatially defined regions enriched in macrophages, adipocytes, and metabolic activity. To validate the main transcriptomic findings of those newly identified subtypes at the protein level, we performed immunohistochemical (IHC) analyses for AR and CD68 on archival frozen samples. Given the semiquantitative nature of IHC evaluation, we used Fisher’s exact test to assess associations between AR IHC score bins and ILC4TME subtypes. AR expression showed a significant association when stratified by quintiles (P = 0.0413), with ARE cases enriched in the top AR expression quintile (consistent trends were also observed using median, tertiles, and quartiles), showing highest values of AR protein expression (Fig. 5E and Dataset S16). CD68 staining, used as a macrophage marker, exhibited a trend toward higher values in the MIE subtype when dichotomized by the median, although this did not reach statistical significance (Dataset S16).

Together, these comparisons reinforce that while there is some conceptual overlap between ILC4TME and existing classifications, our system adds unique dimensions by resolving the cellular and spatial complexity of the ILC microenvironment. In addition to integrating tumor-intrinsic and microenvironmental features that may better inform prognosis and therapeutic targeting, ILC4TME also identifies potentially targetable pathways, offering further opportunities for precision treatment in lobular breast cancer. These transcriptomic findings were further supported by immunohistochemical analyses, which showed corresponding trends in AR and CD68 expression at the protein level. This additional layer of evidence lends support to the biological relevance of the newly defined subtypes and suggests their potential detectability through conventional pathology.

Discussion

Invasive lobular carcinoma, the second most common histological subtype of breast cancer, remains undercharacterized despite its unique clinical and biological features. Current molecular classifications are derived from high-tumor-content bulk RNA sequencing and lack the resolution to capture the spatial organization and microenvironmental complexity of ILC. As a result, identifying clinically relevant subgroups with targetable pathways and distinct outcomes remains challenging.

In this study, we leveraged spatial transcriptomics to profile a large cohort of HR+/HER2− ILC tumors, integrating morphological annotation, single-cell deconvolution, and bulk RNA sequencing for validation in external cohorts. By clustering ST spots based on gene expression profiles, we characterized distinct tumor and TME features, revealing both inter- and intratumor heterogeneity associated with clinical outcomes. Notably, we identified four spatially and biologically defined ILC subtypes: normal/stroma-enriched (NSE), proliferative (P), androgen receptor-enriched (ARE), and metabolic/immune-enriched (MIE). These subtypes were validated in bulk datasets using gene signature-based classification.

A key strength of our work lies in the multiresolution approach: Spatial transcriptomics provided context-specific information on tissue architecture and cell–cell interactions; single-cell data resolved cell-type composition at the spatial level; and bulk RNA-seq enabled validation in large clinical cohorts. This integrative strategy offers a comprehensive and translatable framework for understanding ILC heterogeneity.

The ILC4TME classification retained prognostic value across datasets and clinical settings. Prognostic differences were evident even among patients treated with endocrine therapy alone, where the P subtype was associated with poor outcomes and the NSE subtype with favorable prognosis. These results suggest that the P group may benefit from chemotherapy escalation, while endocrine monotherapy could be sufficient for NSE patients. Importantly, the ILC4TME subtypes maintained prognostic value after adjusting for established gene signatures such as ROR, EndoPredict, MammaPrint, Oncotype DX, Genomic Grade Index, and LobSig, underscoring their added biological relevance. Moreover, although proliferation levels are generally lower in ILC than in NST, we still identified a proliferative subtype of ILC associated with adverse outcomes, suggesting that proliferation retains prognostic significance even at lower levels in this histology, likely reflecting that current prognostic signatures are calibrated on NST biology.

Compared to the previous bulk-derived classification, our subtypes offer higher resolution. While the P and MIE subtypes aligned with the proliferative and immune-related groups of the TCGA classification, respectively, the MIE subtype captured a distinct immune-metabolic niche defined by immunosuppressive M2-like macrophages (31) and adipocytes, suggesting a potential role for adipocytes in macrophage polarization, as described by other studies (32). In addition, comparative analyses with NST tumors demonstrated a strong association between histology and ILC4TME subtype distribution, with MIE and NSE preferentially enriched in ILC and the proliferative (P) subtype more prevalent in NST, while the ARE program showed no strong histology bias. Beyond subtype prevalence, ILC4TME programs displayed higher transcriptional coherence and greater explained variance in ILC than in NST, indicating that these signatures capture more compact and biologically structured programs in lobular carcinoma. Consistent with this, prognostic stratification by ILC4TME revealed clinically meaningful heterogeneity among nonproliferative subtypes specifically in ILC, whereas outcome differences in NST were largely driven by proliferation alone. Exploratory analyses in HER2+ and TNBC ILC further suggest that these programs may extend beyond the HR+/HER2− setting, with enrichment of P in HER2+ and a trend toward enrichment of ARE in TNBC, consistent with known proliferative and androgen-driven biology, although these observations should be interpreted cautiously. The ARE subtype, characterized by androgen signaling and vascularization, represents a category not identified in previous classifications. Expression of AR (also at the protein level) was significantly higher in ARE tumors, reinforcing the biological validity of this group. Moreover, the NSE subtype was associated with better outcomes, with tumor cells more often colocalized with fibroblasts and in situ carcinoma, aligning with prior findings linking in situ presence to less aggressive disease (33). The favorable prognosis associated with the NSE subtype can be at least partly explained by its stromal context. NSE tumors display enrichment of myofibroblastic CAFs (myCAFs), which, despite their canonical tumor-promoting role, have been shown to exert tumor-restraining functions in specific tumor types (34). In ILC, increased myCAF abundance and proximity to tumor cells have been linked with less aggressive disease (35). Moreover, enrichment of apical junction pathways suggests a structurally organized tumor-stroma ecosystem that may physically constrain dissemination and promote epithelial stability, providing a plausible mechanistic basis for improved outcomes in this group (36, 37).

Translating these findings into practice opens promising therapeutic avenues. Tumors classified as metabolic/immune-enriched, characterized by the presence of adipocytes, M2-like macrophages, and activation of metabolic pathways, could be amenable to treatments targeting these features, such as metabolic inhibitors (e.g., Orlistat) (38) or macrophage-modulating agents like PI3K-γ inhibitors (e.g., eganelisib) (39). The androgen receptor-enriched subtype, defined by strong AR expression and related signaling activity, may respond to AR modulators such as enobosarm (40). Consistently, IHC validation on archival frozen samples showed significantly higher AR protein levels in ARE tumors, with similar trends across multiple cutoffs, reinforcing not only the biological validity of this subtype but also its potential for therapeutic targeting with AR-directed agents. CD68 staining showed a parallel, though nonsignificant, trend in MIE tumors, likely reflecting technical limitations and small sample size rather than absence of a biological signal. In contrast, the proliferative subtype, marked by high tumor cell proliferation and associated with worse clinical outcomes, may benefit from chemotherapy escalation strategies, including neoadjuvant or intensified adjuvant regimens. Finally, the normal/stroma-enriched subtype, linked to stromal elements such as CAFs and in situ carcinoma and associated with more favorable outcomes, could potentially be considered for chemotherapy de-escalation. However, while these therapeutic hypotheses are biologically plausible, their clinical applicability remains to be established and will require prospective clinical trials and future functional studies to elucidate the underlying mechanisms and validate their therapeutic potential.

Despite its strengths, the study has some limitations. Although our cohort represents the largest ILC series profiled with spatial transcriptomics to date, it may not capture the full histological diversity of ILC. The relatively small number of spatial transcriptomics samples used to derive the classification is a constraint; however, this was mitigated by extensive validation in large external bulk RNA-seq cohorts. The limited number of relapse events also constrained prognostic analyses within our cohort. Moreover, the deconvolution step relied on a single-cell reference lacking ILC samples. Although this reflects a broader limitation in the field, validation with complementary approaches (e.g., xCell) confirmed the robustness of the inferred cell-type compositions, indicating that the absence of ILC-specific reference data did not compromise the biological conclusions. While we validated the ILC4TME subtypes in external bulk datasets, this approach relies on gene signatures as proxies and lacks direct morphological information. Nonetheless, subtype-specific pathways and cell-type enrichments were consistently observed across cohorts. Minor differences between METABRIC and SCAN-B likely reflect differences in platform (microarray vs. RNA sequencing) and cohort size. These considerations also motivated our focus on SCAN-B for survival analyses: Despite the longer follow-up in METABRIC, SCAN-B provides substantially larger sample size, more contemporary treatment context, and, critically, bulk RNA-seq profiling without tumor cellularity filtering, a prerequisite for faithfully capturing tumor microenvironment and cell type–driven signals. We also observed some differences in clinical associations between discovery and validation cohorts, likely reflecting the limited sample size of the former, while larger validation cohorts provided greater power to detect moderate effect sizes.

In summary, we present a robust, spatially anchored classification of ILC that identifies biologically distinct and clinically relevant subtypes. Our approach expands the current understanding of ILC by integrating tumor-intrinsic features with microenvironmental context, paving the way for more precise prognostic tools and personalized treatment strategies.

Materials and Methods

Spatial transcriptomics (ST) data were generated from frozen tumor tissues of 43 primary HR+/HER2– ILC patients collected at Institut Jules Bordet under approved ethical protocols, with informed consent. Tissue sections were processed according to the Visium Spatial Gene Expression User Guide (10× Genomics) and sequenced on an Illumina NovaSeq 6000 platform at BRIGHTcore. Raw FASTQ files were processed using Space Ranger v1.0.0 with default parameters against GRCh38. Morphological structures were annotated in QuPath v0.2.3, and tumor-stroma features were quantified using an in-house Python script. Downstream analyses, including spot-level processing [STutility (41), Seurat (42)], single-cell deconvolution [CARD (25), xCell (43)], clustering [Louvain algorithm (44)], and enrichment analyses [GSVA (45), fgsea (46)], were performed in R v4.2.1. Multimodal clustering of tumor microenvironment features was implemented with intNMF (27) in the MOVICS (28) package to define four ILC subtypes, further characterized by differential expression [DESeq2 (47)] and gene set analyses. Validation was conducted on METABRIC (20–22) and SCAN-B (23) datasets, with survival analyses performed using the survival (48) and survminer (49) packages. Data and code are publicly available at EGA (EGAD50000001467), Zenodo (https://zenodo.org/records/14924871), Datasets, and GitHub (https://github.com/BCTL-Bordet/ILC-Spatial-Transcriptomics). Detailed Materials and Methods are available in SI Appendix file.

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (XLSX)

Data, Materials, and Software Availability

FASTQ files of the Visium cohort, original code, data and code data have been deposited in ST lobular manuscript dataset on GitHub (https://github.com/BCTL-Bordet/ILC-Spatial-Transcriptomics) (50), EGA (EGAD50000001467) (51), and Zenodo (https://zenodo.org/records/14924871) (52). Anonymized clinical data of ILC patients have been deposited in Dataset S1. Other data are included in the article and/or supporting information.