Heterogeneity of tertiary lymphoid structures predicts the response to neoadjuvant therapy and immune microenvironment characteristics in triple-negative breast cancer

Abstract

Background

Tertiary lymphoid structures (TLSs) impact cancer outcomes, including in triple-negative breast cancer (TNBC), where their role in immune modulation during neoadjuvant therapy (NAT) is underexplored.

Methods

This study employed single-cell RNA sequencing (scRNA-seq), multiplex immunofluorescence (mIF) staining, and radiomic techniques to evaluate TLSs and the tumour microenvironment (TME) in TNBC patient samples before and after NAT.

Results

The presence of TLSs in TNBC was associated with B-cell maturation and T-cell activation. Compared with TLS-low TNBC, TLS-high TNBC showed significantly greater expression of immunoglobulin family genes (IGHM and IGHG1) in B cells and greater cytotoxicity of neoantigen-specific CD8 + T cells (neoTCR8). Additionally, mIF revealed notable differences between TLSs and the TME in TNBC. Although CD8 + T-cell levels do not predict the NAT response effectively, TLS maturity strongly correlated with better NAT outcomes and prognosis (P < 0.05). An imaging biomarker scoring system was also developed to predict TLS status and NAT efficacy.

Conclusion

Our results demonstrated changes in TLSs and the TME in TNBC patients post-NAT. These findings confirm the predictive value of mature TLSs (mTLSs) and support the use of personalised immunotherapy based on post-NAT immune characteristics, thereby improving clinical outcomes.

Introduction

Breast cancer (BC) is now the most common malignancy in women worldwide [1]. Compared with luminal and HER2-positive (HER2 + ) BC, triple-negative breast cancer (TNBC) has the highest recurrence and death rates [2]. For individuals with high-risk early or locally advanced cancer, neoadjuvant therapy (NAT) remains the primary therapy regimen [3]. However, fewer than 30% of TNBC patients achieve complete remission, and their recurrence and death rates are greater than those of non-TNBC patients [4]. Notably, recent research has shown that the infiltration and activation of B or T cells in the TME are related to long-term survival [5]. As a result, investigating the mechanisms regulating immune cell infiltration and activation in the TNBC tumour microenvironment (TME) is worthwhile as these mechanisms may be critical for improving the prognosis of TNBC patients.

Tertiary lymphoid structures (TLSs) are ectopic lymphoid structures that develop in peripheral nonlymphoid organs as a result of chronic inflammation or persistent immunological stimulation and are composed primarily of B and T cells. Increasing data suggest that TLSs play an important role in influencing adaptive antitumour immune responses and are associated with better prognoses in a variety of malignancies, including BC [6, 7]. This observation was confirmed in our previous study [8].

Owing to the lack of oestrogen receptor expression and HER2 amplification, treatment for TNBC largely relies on cytotoxic chemotherapy and immunotherapy. The presence and specific molecular expression patterns of immune cells in the TME may influence the efficacy of immunotherapy [9]. The nature and composition of the TME change over time as a result of chemotherapy. Previous research has suggested that NAT may inhibit the formation of TLSs in malignancies such as non-small cell lung cancer (NSCLC) [10]. Compared with patients who did not receive chemotherapy, those treated with NAT have shown reduced maturation of TLSs and a diminished prognostic value of TLSs. This reduction is thought to be largely due to the immunosuppressive effects of corticosteroids administered alongside chemotherapy [11]. However, the role of chemotherapy in antitumour immunity is still a subject of debate, and the specific influence of NAT on TLS behaviour within the TME remains unclear. Therefore, exploring how TLS characteristics change during NAT is crucial for understanding their impact on clinical outcomes.

Chromosomal instability is a notable hallmark of cancer, primarily manifested through point mutations and copy number variations (CNVs) that drive progressive evolution and enable immune evasion [12]. These chromosomal alterations affect not only the expression of genes in the aberrant chromosomal regions but also related gene expression, potentially correlating with cancer status [13]. Research has demonstrated that significant levels of somatic mutations, common TP53 mutations (83%) and complex aneuploidy rearrangements (80%) contribute to intratumoral heterogeneity (ITH) in patients with TNBC [14]. Recent studies have linked the deletion at 4q21.1 to chemotherapy resistance [15]. However, the relationship between CNVs and TLSs in TNBC remains largely unexplored.

Radiomics, characterised by its noninvasive and cost-efficient attributes, focuses on the quantitative analysis of medical images typically used in standard care. This approach extracts and summarises imaging characteristics such as shape, size, texture, and wavelet decomposition, all of which capture biologically and clinically essential information about tumours. The potential of radiomics to function as predictive and prognostic biomarkers is increasingly recognised by numerous studies. For example, specific radiomic features are able to identify a T-cell presence in lung [16], stomach [17], and liver cancers [18]. A recent investigation introduced a radiomics-derived biomarker that effectively identified CD8 + T-cell infiltration and predicted the efficacy of immunotherapy within a diverse cohort of solid tumours in a clinical trial [19]. Given that TLSs represent a key element of the TME, it is postulated that radiomics could identify the presence of TLSs, allowing their use as a prospective biomarker for NAT.

In this study, we sought to determine the significance of TLSs in providing local antitumour immunity, as well as how distinct CNVs influence TLS production and function in TNBC. We also studied TLSs in paired samples (pre- and post-NAT) from TNBC patients receiving NAT to determine the impact of NAT on TLS abundance and maturation, as well as the relationships between TLS features and pathological response and prognosis. Furthermore, we investigated the use of radiomics as a biomarker for TLSs to facilitate future clinical applications.

Materials and methods

Patients and data

Patients

This study included 63 patients with TNBC who received NAT at Yunnan Cancer Hospital between January 2016 and January 2022 (Supplementary Table 1). In this study, paired formalin-fixed paraffin-embedded (FFPE) tissue samples were obtained from patients before and after NAT, and haematoxylin and eosin (H&E) and multiplex immunofluorescence (mIF) staining were performed on 126 samples. Patients were divided into two groups according to the Miller‒Payne (MP) grading system: response (R) (MP grades 4–5) and nonresponse (NR) (MP grades 1–3). Comprehensive clinical and pathological data were documented for all patients. Disease-free survival (DFS), defined as the time after surgery without symptoms, metastasis, or recurrence, was estimated, with a median follow-up period of 36 months. Additionally, we included an external validation cohort of 45 TNBC patients who underwent NAT at Fujian Cancer Hospital between January 2020 and January 2023 (Supplementary Table 2). All patient samples were collected and handled in strict accordance with the ethical standards of the Declaration of Helsinki. Ethical approval was obtained from the Ethics Committees of Yunnan Provincial Cancer Hospital (Approval No. KYLX2024-060) and Fujian Cancer Hospital (Approval No. K2024-056-01). Informed consent was obtained from all participants prior to inclusion in the study.

Datasets

This study utilised BC cases from The Cancer Genome Atlas (TCGA) database as an initial exploratory cohort. Bulk RNA sequencing data (standardised across all cancers) and gene-level CNVs (derived via Genomic Identification of Significant Targets in Cancer 2 (GISTIC2)) were downloaded for the BRCA cohort from UCSC Xena (http://xena.ucsc.edu/). Additionally, two scRNA-seq datasets (GSE246613 [20] and GSE161529 [21]) were obtained from the Gene Expression Omnibus (GEO). Furthermore, a dataset from our previous study was incorporated to enhance the scientific rigour and breadth of the research [8].

Histopathological and immunohistopathological analysis

To identify and characterise TLSs, we performed histopathological and immunohistochemical analyses on 4-micron continuous paraffin sections. Initially, H&E staining was used to preliminarily identify potential TLS regions on the basis of dense lymphoid cell aggregates. These regions were then confirmed and classified according to their maturation stages via mIF staining for CD20, CD21, and CD23 on adjacent serial sections. Further mIF analyses of additional serial sections were performed to assess the immune cell composition within the TLSs and the surrounding TME. The slides were rehydrated, subjected to antigen retrieval, and blocked overnight at 4 °C with primary antibodies against CD20 (Dako, L26, 1:300), CD21 (Dako, M0784, 1:100), CD23 (Dako, M0823, 1:200), CD4 (Abcam, ab133616, 1:100), CD8 (Abcam, ab178089, 1:100), FoxP3 (Abcam, ab20034, 1:100), PD-L1 (SP142, Roche, 1:100), CD163 (Abcam, ab182422, 1:500), CD68 (Abcam, ab213363, 1:1000), and PD-1 (CST, D4W2J, 86163S, 1:200). After staining, the nuclei were counterstained with DAPI. Visualisation was achieved via SlideViewer and K-Viewer software, and quantitative assessments were performed using QuPath v0.4.1 image analysis software.

TLS assessment

TLSs are classified into three stages of maturity on the basis of the presence of B cells and follicular dendritic cells (FDCs): early TLSs (eTLSs, CD20 + CD21 − CD23 − ), primary follicle-like TLSs (pTLSs, CD20 + CD21 + CD23 − ), and secondary follicle-like TLSs with germinal centres (sTLSs, CD20 + CD21 + CD23 + ). Each sample’s TLS status is determined by the highest stage of maturity present, reflecting the importance of mature TLSs (mTLSs) in the immune response and prognosis. Patients with pTLSs and sTLSs were classified into the mTLS-positive group, whereas those without TLSs or with eTLSs were categorised into the mTLS-negative group. Additionally, TLS abundance is independently scored from 0 to 3 to quantify all TLS stages within the sample, providing a comprehensive evaluation of TLS density and distribution. This scoring involves measuring the number of stained cells per square millimetre and calculating the percentage of positively stained cells among all nucleated cells, averaged over three randomly selected fields. The tumour tissue TLS abundance is scored from 0 to 3: 0 indicates no TLSs; 1 indicates one or two TLSs; 2 indicates at least three TLSs but less than what would warrant a score of 3; and 3 indicates numerous TLSs throughout the tumour area that converge with each other. The assessments were performed by two blinded, experienced pathologists, ensuring objectivity and consistency in the evaluation.

scRNA-seq analysis

This study involved downloading count matrices and metadata for analysis using the Seurat package. The workflow included [1] filtering expression matrices with conditions set for each cell to have fewer than 6000 RNA features and mitochondrial gene proportions under 15% [2]; logarithmically normalising the expression matrix using 10,000 as a scale factor [3]; extracting 3000 genes with the highest variance via the variance stabilising transformation (VST) method [4]; scaling and regressing the expression matrix against covariates, including the RNA count, mitochondrial gene proportion, and cell cycle score [5]; conducting principal component analysis (PCA), followed by cell clustering on the basis of optimised dimensions and resolution [6]; visualising clustering results using UMAP technology [7]; annotating cell clusters using the Single package and manual adjustment on the basis of the most highly expressed marker genes. Detailed subgroup analyses were performed for each cell cluster.

Genomic variant analysis

Somatic mutation and CNV data were downloaded from the Genomic Data Commons (GDC) using the TCGAbiolinks package. Visualisation of the somatic mutation and CNV data (GISTIC output) was conducted using the maftools package. Significant CNV amplifications and deletions were identified using GISTIC 2.0, with set parameter thresholds for amplification or deletion lengths greater than 0.1 and a significance level of P < 0.05.

Predictive model using dce-mri radiomics and machine learning

DCE-MRI images were retrospectively collected from Yunnan Cancer Hospital and Fujian Cancer Hospital prior to the initiation of NAT. All patients underwent initial DCE-MRI scans within two weeks before starting treatment using 1.5- or 3.0-Tesla MRI systems, with detailed information provided in Supplementary File 1. The study design and workflow are depicted in Fig. 7a. We selected the DCE-MRI phase with the highest tumour contrast enhancement for analysis via ITK-SNAP to delineate the region of interest (ROI), aiming to capture the most diagnostically critical information. This phase was deemed most relevant for accurate and meaningful analysis. Two radiologists employed ITK-SNAP software (version 4.0.1) to define the ROIs for primary breast tumours (designated as Label 1). The peritumoral regions were automatically delineated by the Onekey AI platform, categorising extravasation into three zones: 0–2 mm (Label 2), 2–4 mm (Label 3), and 4–5 mm (Label 4). Prior to extracting radiomic features, all segmented ROIs were normalised to standardise the distribution of image voxels using a resampling method. To ensure the reliability of our features, we randomly selected 20 patients for whom two different radiologists independently performed ROI segmentation twice. We calculated the interclass correlation coefficients for the extracted features. Radiomic features, including geometric, intensity, and texture features, were extracted using the PyRadiomics module.

Fig. 7. Predictive model of TLSs using DCE-MRI radiomics and machine learning.

a Flowchart of the machine learning model development process. b Receiver operating characteristic curves for the mTLS predictive model utilising DCE-MRI radiomics and the KNN machine learning algorithm in both the training cohort and the internal validation cohort. c Receiver operating characteristic curves for the mTLS predictive model utilising DCE-MRI radiomics and the KNN machine learning algorithm in the external validation cohort. d Comparison and correlation of the radiomic TLS score in response to neoadjuvant chemotherapy and relationships with TLS abundance, maturation, and mTLS-positive versus mTLS-negative patients in the Yunnan Cancer Hospital (YNCH) cohort. e Comparison and correlation of the radiomic TLS score in response to neoadjuvant chemotherapy and relationships with TLS abundance, maturation, and mTLS-positive versus mTLS-negative patients in the Fujian Cancer Hospital (FJCH) cohort. f DCE-MRI image and mIF staining of a patient with a low radiomic TLS score of 0.226764848. g DCE-MRI image and mIF staining of a patient with a high radiomic TLS score of 0.706994801.

To ensure robustness, we focused on features with high repeatability, retaining those with a Spearman’s rank correlation coefficient greater than 0.9 between any two features. By employing a LASSO regression model on the discovery dataset, we developed a radiomic TLS score by minimising regression coefficients on the basis of the regularisation parameter λ. This method effectively eliminates irrelevant features, allowing us to retain only the most pertinent features for model fitting. Before the machine learning models were constructed, we randomly divided the Yunan Cancer Hospital cohort into a training cohort and an internal validation cohort at an 80:20 ratio. We also included DCE-MRI images obtained prior to NAT from patients at Fujian Cancer Hospital as the external validation cohort. We then applied the K-nearest neighbour (KNN) algorithm to the training cohort to build a predictive model for mTLS based on radiomic features, with mTLS being specifically derived from the results of the aforementioned mIF. The model was then trained to match the labelled data accurately and predict outcomes in the internal validation cohort and the external validation cohort. T-tests or one-way ANOVA were conducted to compare the differences in radiomic TLS scores for mTLS-positive and mTLS-negative cases, as well as across different abundances (scores of 0, 1, 2, and 3), treatment responses (responses vs. nonresponses), and maturities (eTLSs, pTLSs, and sTLSs) in the cohorts from Yunnan Cancer Hospital and Fujian Cancer Hospital.

Statistical analysis

Statistical analyses were performed with R version 4.4.0. Categorical variables were compared with χ² or Fisher’s exact tests. Owing to the low TLS incidence and uneven TLS score distribution, we applied weighted logistic regression to assess the impact of the TLS score on the pathological response. DFS was estimated via the Kaplan‒Meier method and compared with the log-rank test. Univariate and multivariate Cox regression analyses were conducted to evaluate the prognostic significance of the TLSs, with patients stratified based on mTLS status. Wilcoxon tests were used to examine immune cell changes before and after NAT and across different treatment effectiveness levels. Statistical significance was set at p < 0.05.

Results

The impact of TLSs on the prognosis of TNBC

TLSs are considered hotspots of the local immune response, and their presence is associated with improved prognosis and therapeutic efficacy [22]. To systematically dissect the prognostic potential of TLSs and tumour-infiltrating lymphocytes (TILs), we conducted a comprehensive dynamic analysis from the microlevel to the macrolevel and developed predictive biomarkers for TLSs on the basis of radiomic data (Fig. 1f). We used quanTIseq and the 12-CK score [23] to quantify immune cell subgroups and TLSs in the TCGA BC cohort. The correlation of various immune cell subgroups and TLSs with overall survival (OS) was subsequently determined using Cox regression models. The results revealed a significant correlation between the presence of TLSs and B cells with patient OS (Fig. 1a). Notably, xCell analysis revealed that the TLS-high group had a higher Immune Score, Stroma Score, and Microenvironment Score than the TLS-low group (Fig. 1b), suggesting more robust immune activity and a richer stromal composition in the TME, which is often associated with more vigorous tumour immune responses and potential improvements in prognosis [24].

Fig. 1. Impact of TLSs on BC Prognosis.

a Prognostic potential of various immune cell subgroups and TLSs in the TCGA BRCA cohort (n = 1084). Hazard ratios for overall survival are plotted on the x-axis, obtained from univariate (green) and multivariate (blue) Cox regression models. b Immune infiltration scores based on xCell between the TLS-low and TLS-high groups in the TCGA BRCA cohort. c Correlations between B cells and TLS markers in the TCGA BRCA cohort. d Kaplan‒Meier plots for overall survival by TLS status in the TCGA BRCA cohort. e Abundance of TLSs in different subtypes within the TCGA BRCA cohort. f Overview of the study.

Furthermore, a significant correlation was observed between B cells and a TLS presence (r = 0.54, p < 0.0001; Fig. 1c), highlighting the crucial role of B cells in maintaining the TLS structure. Kaplan–Meier analysis of the TCGA BC cohort confirmed that patients with higher TLS scores had better OS (Fig. 1d). Moreover, analysis of TLS abundance in different BC subtypes revealed that some subtypes, particularly TNBC, exhibited higher TLS abundance, potentially due to more intense immune responses. Thus, our subsequent analyses focused on TNBC (Fig. 1e).

These results highlight the importance of a TLS presence in TNBC as a prognostic factor, although the specific biological functions of TLSs in TNBC are still not fully understood.

Functional roles of B cells within TLSs in TNBC

For a deeper understanding of the biological functions of TLSs in TNBC, we conducted a fresh analysis of a recently published scRNA-seq dataset [20]. This dataset provides valuable information on the NAT response in TNBC, with a total of 100,985 immune cells from 24 TNBC patients (Supplementary Fig. 1A). By examining the expression of marker genes, we were able to identify different cell types (B cells, T cells, myeloid cells, and mast cells) among the pooled tumour-infiltrating immune cells from these 24 samples (Supplementary Fig. 1B). After the 12 tumours were categorised into two groups on the basis of B-cell content, namely, TLS-high (n = 12; 63,501 cells) and TLS-low (n = 12; 37,484 cells), we analysed the data (Fig. 2a). Careful analysis of the proportions of each tumour-infiltrating immune cell subtype revealed that TSL-high tumours exhibited a greater abundance of T cells than did TSL-low tumours (Fig. 2b). Furthermore, we found that the NAT response cohort was predominantly associated with TLS-high tumours (Fig. 2c). We also compared the number of cells in each subgroup between TLS-high and TLS-low tumours (Fig. 2d) and introduced the total cell count as a covariate in our statistical model to ensure that downstream analyses were independent of the number of sequenced cells [25]. This allowed us to assess differences between the TLS-high and TLS-low groups adjusted for total cell count.

Fig. 2. Functional roles of B Cells within TLSs in TNBC.

a UMAP reveals the distributions of TLS-high and TLS-low in BC. b Proportions of cellular subtypes in TNBC with TLS-low (n = 12; 37,484 cells) vs. TLS-high (n = 12; 63,501 cells). c Comparison of therapeutic efficacy ratios between TLS-low and TLS-high. d Cell counts of cellular subtypes in TNBC with TLS-low vs. TLS-high. e UMAP annotates subgroups of B cells in BC. f Heatmap of differential gene expression within B-cell subgroups. Red indicates high expression; blue signifies low expression. Dot plots illustrating functional genes in nine B-cell subgroups between TLS-low and TLS-high (g) and between treatment-responsive (R) and nonresponsive (NR) scenarios (h). Expression of IGHM (i) and JGHG1 (j) in B-cell subtypes. p values were determined via the Wilcoxon rank-sum test. ****p < 0.0001; ***p < 0.001; **p < 0.01; *p < 0.05; NS, p > 0.05.

To investigate the differences among B-cell subsets under various TLS conditions, we categorised B cells into nine subsets using unsupervised clustering and visualised them via uniform manifold approximation and projection (UMAP) (Fig. 2e, Supplementary Fig. 1C and 1D). On the basis of the marker genes expressed in each cell cluster (Fig. 2f), we identified these subsets as Bcell_00 (naive), Bcell_01 (IgG+ PC), Bcell_02 (memory B), Bcell_03 (IGLJ1 + PC), Bcell_04 (plasmablast), Bcell_05 (HLA-DRA + PC), Bcell_06 (stressed plasma cell (PC)), Bcell_07 (IgA+IgM+ PC), and Bcell_08 (XBP1 + PC). In all nine B-cell subsets, we noticed more pronounced expression of immunoglobulin family genes, such as IGHM and IGHG1, in TLS-high tumours than in TSL-low tumours (Fig. 2g). Furthermore, the data were organised according to treatment responsiveness (Fig. 2h; Supplementary Fig. 1E), revealing that the expression of immunoglobulin family genes, such as IGHM and IGHG1, was elevated in the response (R) group. These findings indicate that these genes might have a significant effect on treatment outcomes, especially in regulating the body’s immune response to therapy. Afterwards, we performed a thorough analysis of the expression of IGHM and IGHG1 within the cohort. The findings revealed a notable increase in the levels of IGHM and IGHG1 in TLS-high tumours (Figs. 2i, j). These findings suggest that TLSs play a role in increasing the production of antibodies within tumours [26].

Overall, these findings suggest a correlation between TLSs and the maturation and differentiation of B cells into antibody-producing plasma cells, which could enhance localised immune responses against TNBC.

Characteristics of T-cell-mediated inflammation in the TME driven by TLSs

Previous studies have shown that CD8 + T cells have a positive impact on the prognosis of BC patients [27]. Using the same scRNA-seq cohort, we grouped the T cells to investigate whether TLSs enhance their cytotoxic activity (Fig. 3a). Consistent with earlier studies, we identified eight T-cell subsets: Tcell_00 (LYZ + CD4); Tcell_01 (KLRB1 + NK); Tcell_02 (naive T cells); Tcell_03 (regulatory T cells, Tregs); Tcell_04 (GZMK + CD8); Tcell_05 (CD4 + T helper cells, Th); Tcell_06 (LCP2 + CD8); and Tcell_07 (PRF1 + NK) (Fig. 3b). Among these eight T-cell subsets, a significantly greater proportion of T cells was observed in both TLS-high tumours and tumours responsive to therapy (Fig. 3c, d, Supplementary Fig. 1f–i). Previous studies have shown that both the Tcell_04 and Tcell_06 subgroups exhibit the highest expression of neoantigen-reactive CD8 + T cells (“neoTCR8”) and that Tcell_04 is at a terminal stage of T-cell exhaustion [20]. Recent research has revealed that, although predominantly exhausted, neoTCR8s play a critical role in cancer immunotherapy, mediating tumour regression [28].

Fig. 3. Functional T cells within TLSs in TNBC.

a UMAP visualisation of eight T-cell subsets in TNBC. b Heatmap of marker genes across the eight T-cell subsets. Bar graphs depicting the proportional differences in T-cell subsets between the TLS-high and TLS-low groups (c) and between the responsive (R) and nonresponsive (NR) groups (d). Dot plots illustrating the differential expression of functional genes (costimulation, cytotoxicity, and exhaustion) in T-cell subsets between the TLS-high and TLS-low groups (e) and the R and NR groups (f). Expression levels of cytotoxic genes in T-cell subsets within the TLS-high and TLS-low groups (g) and the R and NR groups (h). i Comparison of the proportions of T-cell subsets between TNBC and adjacent nonmalignant tissues. j Comparison of GZMK expression in T-cell subsets between TNBC and adjacent nonmalignant tissues. Differences between groups were assessed via the Wilcoxon rank-sum test. k UMAP visualisation of eight T-cell subsets in TNBC. l Comparison of the proportions of T-cell subsets between the TLS-positive (TLS-p) and TLS-negative (TLS-n) groups. m Comparison of GZMK expression in T-cell subsets between the TLS-p and TLS-n groups. Differences between groups were assessed via the Wilcoxon rank-sum test.

Surprisingly, we observed that both the Tcell_04 and Tcell_06 subsets were predominantly found in the TLS-high group and exhibited increased cytotoxicity (Fig. 3e). Similarly, in the groups that responded to NAT, Tcell_04 and Tcell_06 also showed a similar trend of activity, being significantly more concentrated and more toxic in the R group (Fig. 3f). Furthermore, Tcell_04 and Tcell_06 displayed strong costimulation and exhaustion activities, suggesting that despite their phenotypic exhaustion, they still retain the potential to mediate cytolytic effects.

In a recent study of 29 BC patients, scRNA-seq was used to analyse changes in immune cells before and after anti-PD-1 therapy [29]. Patients whose tumours showed clonal expansion of TILs after treatment had significantly higher frequencies of CD8 + T cells coexpressing CXCL13 and GZMA before treatment than did those without TIL clonal expansion. These findings support the notion that CD8 + T cells coexpressing CXCL13 and GZMA may be reactive cells against tumour neoantigens and play a critical role in anti-PD-1 therapy. Additionally, the expression of CXCL13 and GZMA was examined in the Tcell_04 and Tcell_06 subsets, and both subsets also coexpressed these markers, particularly in the TLS-high and TLS-R groups (Figs. 3g, h). These findings clearly indicate that the cytotoxic T-cell populations in tumours with high levels of TLSs are more active. These findings suggest a stronger immune response in CD8 + T cells in the high-TLS group than in the low-TLS group.

Several studies have documented the existence of bystander T cells in tumours that are not specific to the tumour itself [30, 31]. Using an independent scRNA-seq dataset (n = 19, 24,031 cells) [21], we examined T cells in TNBC and compared them to those in normal breast tissue. Upon reclustering the T cells, we discovered that eight identical T-cell subsets were present in both normal and tumour tissues (Supplementary Fig. 1J). Notably, in TNBC, there was consistently greater enrichment of cytotoxic T cells, namely, Tcell_04 and Tcell_06, than in normal tissue. This observation, supported by Fig. 3i, j and Supplementary Fig. 1K, indicates that these particular cytotoxic T-cell populations play a specific role in TNBC tumours.

Considering the possibility that T cells infiltrating TNBC tumours may come from TLSs or secondary lymphoid organs (SLOs), we speculated that TLSs might increase the local proliferation and infiltration of T cells within the tumour. To test this hypothesis, a study was carried out using mIF to measure the abundance of CD8 + T cells in 22 TNBC tumour tissues. In line with the scRNA-seq data, we observed a notable increase in the quantity of CD8 + T cells in tumours that were TLS-positive, as opposed to TLS-negative tumours (p < 0.05, unpaired t test; Fig. 5g). In our previous research, which included scRNA-seq and immunohistochemical (IHC) staining of TLSs in 14 BC samples [8], we also identified eight T-cell subsets (Fig. 3k) and observed an enrichment of highly cytotoxic Tcell_04 and Tcell_06 in the TLS-positive group (Fig. 3l, m).

Fig. 5. Identification of TLSs in BC.

a H&E staining images of TLSs in TNBC tumour tissues; TLSs are outlined with yellow dashed lines. b Detection rate and distribution of TLS maturity in 63 TNBC patients, with a total of 126 paired tumour samples collected before and after NAT. c Representative mIF images displaying TLSs at various maturation stages: early TLSs (eTLSs, CD20⁺ CD21⁻ CD23⁻), primary follicle-like TLSs (pTLSs, CD20⁺ CD21⁺ CD23⁻), and secondary follicle-like TLSs with germinal centres (sTLSs, CD20⁺ CD21⁺ CD23⁺). d Clinical and pathological characteristics of the 63 TNBC patients. e Comparison of TLS presence and maturity before and after NAT. f Comparison of the TLS area before and after NAT. g Comparison of CD4+ and CD8+ T-cell counts in the TME between mTLS-negative and mTLS-positive groups. h Survival differences between the mTLS-negative and mTLS-positive groups (log-rank test).

These results suggest that TLSs may specifically enhance T-cell activation and proliferation, playing a critical role in mediating the inflammatory TME and thus eliciting optimal antitumour immune responses.

Impact of CNVs on TLS formation in TNBC

In BC, despite the antitumour phenotype exhibited by TLSs, their presence varies significantly among patients [32]. Given that the progression of TNBC is largely driven by CNVs, particularly deletions involving tumour suppressor genes, we hypothesise that genetic alterations in genes related to immune cell communication may occur in TNBC cells, thereby influencing the formation of TLSs. Furthermore, in-depth studies of the relationship between the TME and TLSs have demonstrated that cytokines play crucial roles in the formation of TLSs. These cytokines, including interleukins (ILs), chemokines, and growth factors, not only regulate cell type and activity but also facilitate intercellular communication, thus supporting the formation and maturation of TLSs [33].

On the basis of the aforementioned scRNA-seq data [20], we compared genes involved in cytokine and receptor interactions in TNBC patients with high (TLS-high) and low (TLS-low) levels of TLSs. We discovered significant upregulation of CXCL10 and IL15 in the TLS-high group (Fig. 4a). Using data from our previous study [8], we subsequently constructed a tumour evolutionary tree and observed frequent loss of the chromosome 4q region in the tumour epithelial cells of TLS-negative patients (Fig. 4b). Further analysis of the TCGA TNBC dataset confirmed frequent deletions in the chromosome region harbouring IL15 and CXCL10 (chr4q) in TLS-low tumour samples (Fig. 4c), and this loss was associated with reduced expression of these genes (Fig. 4d, Supplementary Fig. 1L). The percentage of CXCL10 CNVs, which are crucial for T-cell migration to tumours, was as high as 22.27% in BRCA samples (Fig. 4e). We also observed that higher expression of CXCL10 was associated with better survival of TNBC patients (Fig. 4f), likely due to its immunotactic properties. Analysis of the single-cell dataset GSE246613 revealed elevated expression levels of IL15 and CXCL10 in TNBC with substantial infiltration of T cells and myeloid cells (macrophages and dendritic cells) (Fig. 4g, h), which was also validated with TCGA data (Fig. 4i, j). Overall, these findings suggest that the deletion of key cytokines and chemokine genes might be a common mechanism underlying TLS formation in TNBC.

Fig. 4. CNV disruption in TLS formation.

a Heatmap illustrating the differential expression of cytokine signalling genes in tumour cells between the TLS-high and TLS-low groups. b Chromosomal evolutionary trees depict variations in chromosome copy number between the TLS-p and TLS-ne groups. c CNV proportions in the TCGA BRCA cohort show a comparison between TLS-high tumours (in deep red and deep blue) and TLS-low tumours (in light red and light blue), with the genomic locations of CXCL10 and IL15 marked; the x-axis represents chromosome numbers; the y-axis shows the proportion of copy number gains (red, 0 to 1) and losses (blue, 0 to -1). The TLS signature is derived from the 12-CK score. d Correlations between the copy number status of CXCL10 and its mRNA expression levels in the TCGA cohort, with Spearman and Pearson correlation coefficients provided. e The number of tumours with CXCL10 deletion across different cancer types; the x-axis indicates the number of cases; green represents cases with partial CXCL10 deletions; grey denotes cases without CXCL10 deletions; pink indicates amplified cases. f Kaplan‒Meier analysis of CXCL10 expression levels in TNBC patients from the TCGA BRCA cohort. g Dot plot displaying the expression of cytokines and their receptors in cell subtypes within the TLS-high and TLS-low groups. h Comparison of IL15 and CXCL10 expression in different cell types from TLS-high and TLS-low tumours in the GSE246613 dataset. i, j Scatterplots showing the correlation between CXCL10 expression levels and the infiltration of M1 macrophages and dendritic cells in TNBC.

Identification of tumour-Intrinsic TLS in TNBC

To identify TLSs and decipher their immunodynamics during NAT for TNBC, we analysed pre- and posttreatment tumour tissue samples from 63 TNBC patients using H&E and mIF staining. The clinicopathological characteristics of the TNBC patients included in this study are presented in Fig. 5d and Supplementary Table 1.

The presence of tumour-intrinsic TLSs was initially assessed via H&E staining, which was used to determine the proportion and size of TLSs in each individual. There was significant heterogeneity in the shape and area of the TLSs, with most appearing oval or irregular (Fig. 5a). A broader overview of the H&E-stained slides is provided in Supplementary Fig. 2A. For the assessment of intratumoral TLS maturation stages, mIF staining was conducted. Using mIF, we examined the presence of B cells (CD20 marker), follicular dendritic cells (CD21 marker), and germinal centre (GC) B cells (CD23 marker) within the TLSs. On the basis of the expression of these markers, we classified the TLSs into three stages of maturation: eTLSs, pTLSs, and sTLSs (Fig. 5b, c; Supplementary Fig. 2B).

In this study, we included 63 TNBC patients and collected paired tumour samples before and after NAT, totalling 126 samples. Among these samples, 35 were found to contain TLSs, with 23 exhibiting high maturity (Fig. 5b). Among the pre-NAT samples, TLS positivity was observed in 24 patients (38.10%), with 18 patients (28.57%) exhibiting high levels of mTLSs. In contrast, among the post-NAT samples, only 11 patients (17.46%) were TLS positive, and high levels of mTLSs were present in only 5 patients (7.94%) (Fig. 5e). These findings suggest that NAT may impact both the abundance and maturity of TLSs. Moreover, the size of the intratumoral TLSs after NAT was significantly smaller than that before NAT (Fig. 5f).

In addition, TNBC samples that exhibited intratumoral mTLS structures presented notably elevated levels of CD8 + T-cell and CD4 + T-cell infiltration within the TME than did TNBC samples without mTLSs (Fig. 5g). There seems to be a strong connection between the development of TLSs and the infiltration of immune cells, implying that they have a functional relationship.

Association of pre-treatment TLSs with improved NAT response and prognosis in TNBC

We explored the potential correlation between the maturity of TLSs in pretreatment samples and known clinicopathological features in TNBC patients. A statistically significant association was found only between TLSs and the T stage of the tumour (p < 0.05), with patients with low-maturity TLSs presenting at a later T stage than those with high-maturity TLSs. No significant correlations were detected between TLS maturity and other clinical characteristics (Table 1).

Table 1.

The correlation between TLS maturation and clinicopathological features

Characteristics		Low-maturation	High-maturation	P
Age	<50	28	11	0.935
	≥50	17	7
Menopause	Yes	20	5	0.222
	No	25	13
T stage	T1-T2	21	14	0.025
	T3-T4	24	4
Lymphnode metastasis	Yes	38	13	0.264
	No	7	5
TNM stage	II	21	12	0.151
	III	24	6
Ki67	<20%	3	2	0.555
	≥20%	42	16

Compared with those in the mTLS-negative group, individuals in the mTLS-positive group were much more likely to achieve a response (p < 0.05, Fig. 6f). The analysis revealed that the presence of mTLSs in pre-NAT biopsies was linked to a more favourable response to NAT. The univariate odds ratio was 7.00, with a 95% CI of 1.44–34.06, and the p value was less than 0.05 (Table 2). Through careful analysis, we examined the relationships of TLS maturity and abundance with DFS in both groups. During the 36-month follow-up period, 15 patients in the cohort of 63 NAT patients experienced disease recurrence or progression. According to the Kaplan‒Meier analysis, the group with higher maturity had a longer DFS than did the group with lower maturity (p < 0.05, Fig. 5h). Univariate and multivariate Cox regression analyses, stratified by mTLS status, revealed that in the mTLS-negative group, T stage and lymph node metastasis were significant independent predictors of DFS (Table 3). In contrast, in the mTLS-positive group, these traditional prognostic factors were not significantly associated with DFS (Table 4). These findings suggest that TLSs may improve patient prognosis.

Fig. 6. Assessing the heterogeneity of TLS and TME using mIF.

a–d Using two sets of specific antibodies, the maturity of TLSs in 22 TNBC tissue samples was analysed before and after NAT, along with the heterogeneity of the TME (c, d) through mIF. (e) Comparison of the primary immune cell counts in the TME before and after NAT treatment. (f) Analysis of mTLS-positive and mTLS-negative proportions in patient groups categorised as nonresponders (NRs) and responders (Rs) to NAT. (g) Comparison of primary immune cell counts in the TME between the NR and R groups.

Table 2.

Univariate analysis of response to NAT

Characteristics		Non-Response (n = 23)	Respnse (n = 40)	P	OR (95%CI)
Age	<50	16	23	0.344	1.69 (0.57–5.01)
	≥50	7	17
Menopause	No	15	23	0.547	1.39 (0.48–4.01)
	Yes	8	17
T stage	T1-T2	18	31	0.944	1.05 (0.30–3.60)
	T3-T4	5	9
Lymphnode metastasis	No	2	10	0.129	0.29 (0.06–1.44)
	Yes	21	30
TNM	II	9	24	0.114	0.43 (0.15–1.22)
	III	14	16
Ki67	<20%	2	3	0.866	1.17 (0.18–7.60)
	≥20%	21	37
NAT number of cycles	6	9	9	0.164	2.21 (0.72–6.78)
	8	14	31
TLS maturation	Low-maturation	21	24	0.016	7.00 (1.44–34.06)
	High-maturation	2	16
TLS abundance	Score 0	18	22	0.92	1.22 (0.24–62.83)
	Score 1	2	6	0.305	2.45 (0.44–13.67)
	Score 2	2	9	0.122	3.68 (0.70–19.25)
	Score 3	1	3	0.453	2.45 (0.23–25.67)

Table 3.

Univariate and multivariate analyses of DFS following NAT in the mTLS-negative Group

Characteristics		Univariate analysis		Multivarible analysis
		HR (95%)	P	HR (95%)	P
Age	<50 vs. ≥50	0.851 (0.210–3.459)	0.822
Menopause	Yes vs. No	0.314 (0.050– 1.303)	0.052
T stage	T1-T2 vs. T3-T4	0.054 (0.008–0.371)	0.003	0.105 (0.016–0.696)	0.02
Lymphnode metastasis	Yes vs. No	0.153 (0.012–0.972)	0.042	0.232 (0.028–0.935)	0.037
TNM stage	III vs. II	42.491 (3.192–565.638)	0.005	0.232 (0.028–1.936)	0.008
Ki67	<20% vs. ≥20%	0.071 (0.004–1.409)	0.083
NAT number of cycles	6 vs. 8	0.577 (0.139–2.402)	0.45

Table 4.

Univariate analysis of DFS following NAT in the mTLS-positive group

Characteristics		Univariate analysis
		HR (95%)	P
Age	<50 vs. ≥50	1.333 (0.009–24.994)	0.864
Menopause	Yes vs. No	1.333 (0.009–24.994)	0.864
T stage	T1-T2 vs. T3-T4	1.333 (0.009–24.994)	0.864
Lymphnode metastasis	Yes vs. No	1.286 (0.069–187.612)	0.875
TNM stage	III vs. II	4.5 (0.240–656.641)	0.315
Ki67	<20% vs. ≥20%	0.083 (0.0005–1.562)	0.094
NAT number of cycles	6 vs. 8	1.037 (0.032–1.694)	0.069

Assessing the heterogeneity of TNBC TLS maturation and cellular components through mIF

Understanding the diverse maturation stages of TLSs and their impact on TNBC prognosis highlights the importance of delving deeper into TLS heterogeneity. This includes studying the cellular composition and maturation and how they affect antitumour immune responses. In this study, we randomly selected 11 patients with pre- and post-NAT TNBC and utilised serial tissue sections to thoroughly examine the development and cellular makeup of TLSs.

By employing two different antibody panels, we were able to assess the cellular heterogeneity within the TME comprehensively. Specifically, Panel 1 (CD8, CD4, FoxP3, PD-1, and DAPI) and Panel 2 (CD68, PD-L1, CD163, and DAPI) allowed us to analyse the infiltration of various immune cells across different TLS maturation stages.

Our analyses revealed significant heterogeneity in TILs across various stages of TLS maturation. In contrast to our initial hypothesis, the infiltration of CD8⁺ T cells did not increase with TLS maturity; instead, eTLSs and sTLSs presented greater numbers of CD8⁺ T cells than did pTLSs (Fig. 6a). Conversely, we observed a clear decrease in M2 macrophage (CD68⁺CD163⁺) infiltration with increasing TLS maturity (Fig. 6b), further highlighting the complexity of TLSs and their immune microenvironment.

To explore the potential mechanisms of NAT, we compared the TME before and after NAT treatment, as well as among different response outcomes (Fig. 6c, d). Our analysis revealed that following treatment, there was a significant increase in the number of CD8⁺ T cells within the tumour stroma, while the numbers of FoxP3⁺ T cells, FoxP3⁺ CD4⁺ regulatory T cells (T-regs), CD68⁺ CD163^- M1 macrophages, and CD68⁺CD163⁺ M2 macrophages decreased significantly (p < 0.05, Fig. 6e). However, the numbers of PD-L1⁺ and PD-1⁺ cells did not change significantly after treatment (p = 0.671, Fig. 6e), highlighting the potential role of NAT in enhancing the tumour-rejecting immune environment.

Additionally, we observed an intriguing phenomenon: after NAT, the number and maturity of TLSs decreased (Fig. 5e), whereas the number of CD8⁺ T cells increased significantly (Fig. 6e). To further validate our findings, we included an external validation cohort (Supplementary Table 2). This external cohort exhibited trends similar to those in our initial study—specifically, post-NAT, TLS abundance and maturity decreased, whereas CD8⁺ T-cell numbers increased significantly (Supplementary Fig. 3A–3B). We speculate that the reduction in TLSs after NAT may reflect a shift in the immune response from an initial activation phase to one dominated by effector CD8⁺ T cells, potentially representing a rebalancing of the immune system induced by NAT.

Furthermore, we explored the characteristics of TLSs and immune cell infiltration within the microenvironment across different therapeutic outcomes. We found a significant association between TLS maturity and treatment response. Specifically, patients responsive to NAT exhibited more mTLS states (p < 0.05, Fig. 6f). Surprisingly, there was no significant difference in CD8 + T-cell infiltration between the responsive and nonresponsive groups (p > 0.05, Fig. 6g); differences in immune cells were evident only in FoxP3 + CD4 + T-regs, CD68 + CD163- M1 macrophages, and CD68 + CD163 + M2 macrophages (p < 0.05, Fig. 6g).

These findings suggest that the efficacy of NAT may be closely linked to a more active immune microenvironment, potentially facilitated by the activation of TLSs.

Predictive Model Using DCE-MRI Radiomics and Machine Learning

A total of 428 radiomic features were extracted from DCE-MRI images. Using the LASSO logistic regression model, we developed the radiomics TLS score by selecting features with nonzero coefficients. Supplementary Fig. 3C and 3D illustrate the coefficients and mean squared error (MSE) from the 10-fold cross-validation. The final nonzero coefficients of the selected features are shown in Supplementary Fig. 3F. All of the chosen features demonstrated low correlations with each other, with correlation coefficients ranging from −0.675 to 0.773 (Supplementary Fig. 3E). The radiomics TLS score formula is as follows:

Radiomics TLS score = 0.2857 − 0.0094×original_firstorder_Skewness_L1 − 0.0130×original_gldm_DependenceEntropy_L1 + 0.0182×original_gldm_SmallDependenceLowGrayLevelEmphasis_L2 − 0.0101×original_firstorder_Skewness_L4 + 0.0458×original_glcm_InverseVariance_L3 + 0.0344×original_gldm_SmallDependenceLowGrayLevelEmphasis_L3 − 0.0263×original_glszm_LargeAreaHighGrayLevelEmphasis_L3 + 0.0805×original_glszm_SmallAreaLowGrayLevelEmphasis_L3.

The radiomics TLS score was significantly correlated with the presence of mTLSs and tended to increase with increasing TLS abundance scores in both the Yunnan Cancer Hospital cohort and Fujian Cancer Hospital cohort (Figs. 7d, e). Additionally, it was significantly associated with maturation and therapy response, which is consistent with the known relationship between TLSs and response to NAT in both cohorts (Figs. 7d, e). Furthermore, the selected radiomic features were utilised to construct a KNN-based radiomic model. The ROC curve indicated excellent predictive performance in both the internal validation cohort (AUC: 0.852, 95% CI: 0.624–1.000; Fig. 7b) and the external validation cohort (AUC: 0.874, 95% CI: 0.758-0.990; Fig. 7c). We also present MR images and mIF staining from two patients with low or high radiomics TLS scores. The first patient, with a low radiomics TLS score of 0.226764848, underwent a regimen of four cycles of EC (epirubicin + cyclophosphamide) combined with four cycles of T (albumin-bound paclitaxel) (Fig. 7f). mIF staining did not reveal the presence of mTLSs (Fig. 7f), and this patient did not achieve pathological remission after eight cycles of treatment, resulting in a poor outcome. In contrast, another patient with a high radiomics TLS score of 0.706994801 received the same NAT regimen (Fig. 7g). mIF revealed the presence of mTLSs (Fig. 7g), and this patient achieved a pathological complete response after eight cycles of treatment, resulting in a favourable outcome. These findings suggest that the radiomics TLS score is correlated with the presence of mTLSs and has strong predictive performance for the response to NAT.

Discussion

In this study, we conducted a comprehensive multidimensional analysis of TLSs in paired TNBC samples before and after NAT by integrating single-cell transcriptomics, multiplex imaging, and radiomics. Initially, we explored the impact of TLSs on the tumour ecosystem at single-cell resolution and revealed that TLSs are associated with the activation of B cells and T cells within the tumour. Further investigations into the potential effects of NAT on TLSs and the TME in TNBC patients suggested that mTLSs may serve as a novel biomarker for a NAT response. Finally, we also developed a radiomics signature to identify the presence of TLSs. These findings enhance our understanding of the immunomodulatory effects of NAT and could aid in the development of personalised treatment approaches for TNBC.

B cells constitute the primary immune cell population within TLSs. Our findings indicate that, among patients who underwent NAT, tumour-infiltrating B lymphocytes exhibit variations in distribution and function between the response and nonresponse groups. Furthermore, the presence of TLSs can more effectively differentiate these two scenarios. These findings suggest that B cells that aggregate in TLSs may play an antitumour role and that the microenvironment provided by a TLS facilitates B-cell maturation. This discovery is consistent with our previous research [8]. However, the underlying mechanisms require further exploration.

In fact, clinical trials in TNBC have demonstrated that a combination of immune checkpoint blockade (ICB) with standardised chemotherapy offers a significant therapeutic advantage over ICB monotherapy. This combination strategy has been shown to increase the rate of complete response and prolong DFS [34]. Additionally, the Keynote-522 study series presented at the 2023 San Antonio Breast Cancer Symposium, particularly LBO1-01, strongly indicated that ICB therapy achieves substantial success in TNBC [35]. A phase II trial by Leonie et al. revealed that administering nivolumab after chemotherapy or radiotherapy in patients with metastatic TNBC resulted in high response rates and durable therapeutic effects [36]. Recent studies further confirmed the benefits of sequential combination strategies in animal models, showing that blocking PD-1 postchemotherapy, rather than concurrently, more effectively maintains the proliferation of CD8 + T cells and their antigen-specific cytotoxicity [37]. Consistent with these findings, our comparison of paired TNBC samples before and after NAT revealed an increase in PD-1 expression on CD8 + T cells, regardless of treatment sensitivity or resistance (Supplementary Fig. 2C), providing a plausible explanation for why TNBCs with a “cold” immune microenvironment significantly benefit from the combined use of ICB drugs and chemotherapy.

Consistent with prior studies, our findings indicate that overall CD8 + T-cell levels are not predictive of NAT response in TNBC patients. However, we observed greater enrichment of highly cytotoxic T-cell populations in TLS-high tumours than in TLS-low tumours or normal tissues. These cytotoxic T-cell groups resemble the recently identified tumour-infiltrating neoTCR8s, which are typically exhausted and significantly coexpressed with CXCL13 and GZMA, and are associated with favourable tumour responses [38]. Notably, compared with TLS-low tumours, TLS-high tumours presented significantly greater neoTCR8 activity and numbers. This increase led to increased cytotoxicity and exhaustion, emphasising the important role of TLSs in activating cytotoxic T cells that attack tumour cells. In addition, the creation of TLSs involves multiple steps. Similar to findings in previous studies [39], our research also revealed the common occurrence of deletions in the T-cell regulators IL15 and CXCL10 in TNBC. These deletions often coincide with the loss of TLSs in TNBC. CXCL10 has been identified as one of the twelve cytokine gene signatures that have the potential to indicate the presence of TLSs and is correlated with patient survival in patients with colorectal cancer [23]. While deletions of IL15 and CXCL10 were consistently observed in the TNBC cohort in this study, additional research is needed to determine the direct functional impact of these genes on TLS formation.

The nature and composition of the TME dynamically change under the influence of NAT [40, 41]. The specific impact of NAT on the formation of TLSs in TNBC remains unclear. In our study, paired H&E and mIF analyses conducted before and after NAT revealed a significant reduction in both the maturity and area of the TLSs. In a related study on lung cancer, NAT had a negative impact on the prognostic value of TLS density. This could be attributed to the hindered formation of germinal centres within TLSs [42]. However, other studies suggest that NAT may induce the release of new antigens and immune activation in tumours, thus offering potential benefits [43]. These findings indicate that NAT may have both beneficial and detrimental effects on the formation and development of TLSs. Understanding the intricate relationship between NAT and antitumour immunity is crucial when considering its overall impact on patients receiving this treatment. In addition, we examined the relationships between TLS maturity and clinicopathological characteristics. These findings revealed that the maturity of TLSs was correlated with tumour size. This implies that individuals with lower TLS maturity tended to have a more advanced T stage than those with higher TLS maturity.

Our study confirmed that the protective effects of TLSs vary depending on the type of cancer or clinicopathological parameters, such as tumour stage [44, 45], as previously investigated in various cancers. We investigated the correlation between TLSs and clinical outcomes in patients with TNBC who underwent NAT. Our research suggests that the presence of mTLSs in tumours has significant implications for patients and is a reliable indicator of treatment response and DFS in TNBC patients. In addition, there was noticeable variation in the maturity of TLSs within tumours, possibly indicating variations in their effectiveness against tumours. In regard to measuring TLSs, their level of maturity seems to hold more predictive value for different types of cancers. Through the categorisation of patients into low- and high-maturity TLS groups, it was observed that individuals in the high-maturity group had improved DFS. Therefore, the maturity of TLSs is a strong indicator of positive outcome in patients receiving NAT, potentially surpassing the importance of TLS density in effectively predicting the response of TNBC patients undergoing NAT.

Additionally, we examined paired tumour samples before and after NAT, focusing on features of the TME. Interestingly, we observed a decrease in the number and maturity of TLSs post-NAT, whereas CD8⁺ T-cell counts increased significantly. This paradox may be explained by the role of TLSs in priming CD8⁺ T cells prior to NAT, as mTLSs facilitate initial T-cell activation, leading to the accumulation of activated CD8⁺ T cells in the TME. Despite NAT disrupting TLS structures, these activated CD8⁺ T cells may persist and continue exerting antitumour effects independent of TLSs. Moreover, NAT may remodel the TME by reducing the number of immunosuppressive cells, such as FOXP3⁺ regulatory T cells and M1/M2 macrophages, relieving the suppression of CD8⁺ T cells and enhancing their function [46]. To validate our findings, we included an external validation cohort of 45 TNBC patients who underwent NAT. This cohort exhibited similar trends—decreased TLS abundance and maturity post-NAT, coupled with increased CD8⁺ T-cell counts. These findings suggest that the reduction in TLSs may reflect a shift from an initial immune activation phase to one dominated by effector CD8⁺ T cells, indicating NAT-induced immune rebalancing. Even with reduced TLSs, CD8⁺ T-cell-mediated antitumour immunity can be enhanced, highlighting the potential for combining NAT with immunotherapy. However, our study’s limitations include an insufficient sample size and complexity of TME interactions, necessitating further research to elucidate the underlying mechanisms and optimise treatment strategies.

In the context of NAT, the role of TLSs has been explored in various cancers. Costa et al. demonstrated that in unmatched pancreatic cancer samples, the quantity of TLSs correlates with patient immune cell infiltration and the anti-TME [47]. Our research, by analysing paired samples from TNBC patients, revealed a direct link between TLS maturity and treatment efficacy and prognosis, emphasising the importance of assessing TLS maturity in patients with breast cancer, particularly TNBC. Consistent with the findings of Lu et al. [48], we found that B cells play a crucial role in the immune response induced by chemotherapy. However, our study further explored the heterogeneity of B cells within TLSs and their function in TNBC and revealed that mTLSs with B cells are particularly critical for activating local immune responses. This not only complements Lu et al.’s view on the multifunctionality of B cells but also underscores the clinical relevance of assessing TLS maturity in TNBC. This will provide new directions for future research and clinical applications, particularly in the context of personalised and precision medicine.

By transforming traditional medical imaging data into mineable quantitative data, radiomics can be used to assess the internal characteristics of tumours, and multiple potential predictive and prognostic biomarkers have been developed in recent years [49, 50]. For this study, we used standardised DCE-MRI data to create a radiomic scoring system for TLSs. Using these radiomic data, we were able to accurately identify the presence of TLSs in TNBC patients. Furthermore, the TLS biomarker derived from radiomics has the ability to predict the response of TNBC patients to NAT. This highlights the promising prospect of using radiomic biomarkers to categorise patients for immunotherapy in the future. One thing to consider is that the current TNBC NAT cohort used in this study is relatively small. To confirm these findings, it would be beneficial to conduct further validation in larger clinical trial cohorts.

Conclusion

In conclusion, we elucidated the role of TLSs in promoting the maturation of B cells and T cells, thereby activating local antitumour immune responses in TNBC. Additionally, our study documented the dynamic pathologic landscape of TLSs in TNBC patients before and after NAT. We identified TLS maturity as a biomarker that improves the accuracy of predicting DFS in TNBC patients undergoing NAT. Finally, we developed a biomarker scoring system to identify the status of TLSs and predict the efficacy of NAT, which is poised for clinical translation.

Author contributions

CGS, SCT and ZHL supported and designed the study. QW, XWH, XFL, YS, and JLL collected, processed, and analysed the datasets. CXW, ZRJ, and XWH verified the data. QW and YSY wrote the manuscript. All of the authors have read and approved the final version of the manuscript.

Funding

The present study was supported by the High-level Talent Introduction Project of Fujian Cancer Hospital (No.: F2328R-GC301-01) and the High-level Talent Training Program of Fujian Cancer Hospital (No.: 2024YNG03).

Data availability

The data are available upon reasonable request. All data relevant to this study are included in the article or uploaded as online supplemental information.

Code availability

The underlying code for this study, along with the associated training and validation datasets, is not publicly available. However, it can be made available to qualified researchers upon a reasonable request to the corresponding author.

Competing interests

The authors declare no competing interests.

Ethical approval

This study was approved by the Ethics Committee of Yunnan Provincial Cancer Hospital (Approval No. KYLX2024-060) and Fujian Cancer Hospital (Approval No. K2024-056-01), and written informed consent was obtained from all participants.

Supplementary information

The online version contains supplementary material available at 10.1038/s41416-024-02917-y.