PFAS is associated with perineural invasion in triple-negative breast cancer with a potential role for Cathepsin D dysregulation: a multi-omics and experimental study

Abstract

Per- and polyfluoroalkyl substances (PFAS) are persistent pollutants linked to breast cancer (BC), but their role in perineural invasion (PNI) of triple-negative breast cancer (TNBC) is unclear. Cathepsin D (CTSD), a lysosomal protease, is hypothesized to mediate PFAS-induced PNI, though systematic evidence is lacking. We integrated multi-omics data from TCGA-BRCA, METABRIC, and single-cell RNA-seq datasets. Analyses included differential gene expression, Mendelian randomization, consensus clustering, and machine learning for prognostic modeling. Single-cell analyses were performed using Seurat, Monocle2, and CellChat. GraphBan screened natural CTSD-binding compounds, with binding affinity evaluated by molecular docking and dynamics simulations. Experimental validation included immunohistochemistry, immunofluorescence, Transwell, and Western blot assays. We identified 5 PFAS-associated PNI-related genes (PPGs), with CTSD central to TNBC PNI. PPG-based molecular subtyping revealed a high-risk subgroup exhibiting enhanced epithelial-mesenchymal transition (EMT) activity, proliferation capacity, and significantly poorer overall survival. The PPG-based prognostic model effectively stratified patient outcomes and immunotherapy response. Mendelian randomization confirmed a causal link between genetically predicted CTSD levels and BC risk. Single-cell analysis showed CTSD specifically enriched in myeloid cells; CTSD⁺ myeloid cells displayed immunosuppressive signatures and therapy resistance. CTSD⁺ epithelial cells interacted with cancer-associated fibroblasts via FGF signaling and showed altered metabolism. GraphBan predicted and experiments confirmed Aurantio-obtusin as a high-affinity CTSD inhibitor. Molecular simulations demonstrated stable binding of both PFAS and Aurantio-obtusin to CTSD. Histologically, elevated CTSD expression co-localized with CD68⁺ macrophages in PNI-positive TNBC tissues, while Aurantio-obtusin suppressed CTSD expression and inhibited TNBC cell proliferation and migration. This study suggests that PFAS exposure is associated with PNI and malignant progression in TNBC, potentially involving dysregulation of CTSD. The robust PPG-based prognostic signature and the natural inhibitor Aurantio-obtusin offer novel biomarkers and a potential therapeutic strategy for mitigating PFAS-related cancer risks.

Supplementary Information

The online version contains supplementary material available at 10.1007/s10238-026-02164-w.

Background

Breast cancer (BC) remains the most frequently diagnosed cancer and a leading cause of cancer-related mortality in women worldwide, with approximately 778,000 annual deaths [1]. Triple-negative breast cancer (TNBC), characterized by the absence of estrogen receptor, progesterone receptor, and HER2 expression, represents the most aggressive subtype with the poorest prognosis [2]. Metastasis is associated with TNBC mortality, affecting over one-third of patients [3]. Beyond conventional dissemination routes, perineural invasion (PNI) has emerged as a "third metastatic pathway," strongly associated with tumor aggressiveness and poor outcomes [4, 5]. While immune checkpoint blockers (ICBs) have transformed TNBC management, their efficacy remains limited by PD-1 resistance [6, 7]. Emerging evidence suggests that cancer-induced neural remodeling may contribute to anti-PD-1 resistance [8], positioning PNI as a potential modulator of immunotherapy failure. Therefore, elucidating the molecular mechanisms underlying PNI is crucial for devising strategies to overcome treatment resistance and improve survival in TNBC patients.

Beyond tumor-intrinsic factors, exogenous environmental exposures are increasingly implicated in cancer pathogenesis. Per- and polyfluoroalkyl substances (PFAS), a class of widespread and persistent environmental contaminants, pose significant public health risks due to their bioaccumulative potential and associated health risks [9–11]. Epidemiological studies have indicated a positive correlation between serum PFAS levels and BC incidence [12, 13]. Experimental evidence demonstrates that PFAS, such as perfluorooctanoic acid (PFOA), can promote tumor invasion and metastasis in various cancers by modulating cellular mechanical properties, cell adhesion, and immune evasion [14, 15]. In BC, PFAS has been shown to drive cell proliferation and migration [16]. However, the specific role and mechanisms through which PFAS regulates highly aggressive phenotypes like PNI in BC remain largely unexplored.

Cathepsin D (CTSD), a key lysosomal aspartic protease, is involved in intracellular degradation and metabolic homeostasis [17, 18], and plays a central role in cell fate decisions [19]. CTSD is overexpressed in various malignancies, including BC, where it serves as an independent predictor of early recurrence and poor prognosis [20]. Functionally, CTSD knockout can delay tumorigenesis by suppressing mTORC1 signaling [21]. Furthermore, CTSD is deeply involved in tumor microenvironment (TME) remodeling and therapy resistance; its expression is associated with chemotherapy resistance [22], while its inhibition can promote macrophage repolarization toward the M1 phenotype, thereby suppressing epithelial-mesenchymal transition (EMT) and metastasis [23]. Antibody-based targeting of CTSD has been shown to exert anti-tumor effects by reshaping the immune microenvironment [24]. Notably, environmental pollutants such as nanoplastics and arsenic can disrupt autophagic flux or lysosome-mediated immune surveillance by interfering with CTSD function, thereby promoting tumor progression [25–27]. These findings strongly suggest that CTSD acts as a critical molecular node mediating the carcinogenic effects of pollutants. Additionally, overexpression of CTSD contributes to PNI of salivary adenoid cystic carcinoma [28]. Nevertheless, a systematic investigation into whether and how CTSD mediates the tumor-promoting effects of PFAS in BC, particularly in TNBC and its PNI phenotype, is warranted.

To address these knowledge gaps, we investigated whether PFAS exposure is associated with perineural invasion and malignant progression in TNBC, and whether this association involves dysregulation of CTSD. To test this hypothesis, we integrated multi-omics data from TCGA-BRCA, METABRIC, and single-cell RNA-seq datasets (GSE161529, GSE266919) with Mendelian randomization, computational toxicology, and experimental validation. Our study aimed to: (1) identify PFAS-associated PNI-related genes and construct a prognostic signature; (2) establish CTSD as a core mediator linking PFAS exposure to TNBC risk using Mendelian randomization and single-cell transcriptomics; (3) characterize the role of CTSD + cells in promoting malignant progression through metabolic reprogramming, immunosuppression, and intercellular communication; and (4) screen and validate natural CTSD-inhibiting compounds through molecular docking, dynamics simulations, and cellular experiments. This work delineates a novel environmental exposure-driven pathway in TNBC progression and provides insights for biomarker discovery and targeted therapeutic intervention.

Materials and methods

Dataset source and preprocessing

Publicly available BC transcriptomic datasets—including METABRIC and TCGA-BRCA were downloaded from TCGA (via UCSC Xena, https://gdc.xenahubs.net/) and cBioPortal (https://www.cbioportal.org/). TNBC cases were identified as follows: in TCGA-BRCA, samples with negative status for ER, PR, and HER2 were selected based on clinical receptor annotations. In METABRIC, basal-like subtypes were defined as TNBC according to the PAM50 intrinsic subtype classification. Single-cell RNA-seq data of TNBC were obtained from the GEO repository under accessions GSE161529 and GSE266919. The IMvigor210 immunotherapy cohort was sourced from the IMvigor210CoreBiologies R package (http://research-pub.gene.com/IMvigor210CoreBiologies/). The 2D and 3D structures of PFOA and PFOS were retrieved from PubChem. Potential protein targets of PFAS were collected from a published study [29] and expanded using SwissTargetPrediction (http://swisstargetprediction.ch/). All target identifiers were standardized via the UniProt database (https://www.uniprot.org/).

Identification and functional analysis of perineural invasion-associated targets in BC

A total of 101 perineural invasion-related genes (PNIGs) were collected from a previously published study [30]. While these genes were originally identified in gastric cancer, they have been functionally implicated in PNI biology across multiple cancer types. We acknowledge that direct PNI annotation is not available in the TCGA-BRCA and METABRIC cohorts; therefore, our analyses using these genes serve to identify transcriptomic signatures associated with PNI potential rather than directly confirmed PNI status. Differential expression analysis between BRCA and normal breast tissues was performed using DESeq2, with genes meeting |log₂FC|≥ 1 and adjusted p-value < 0.05 defined as differentially expressed genes (DEGs). The intersection between upregulated DEGs and PNIGs was identified as perineural invasion-associated genes in BC (PGs). Functional enrichment analysis of PGs was conducted using Metascape (https://metascape.org) for GO and KEGG pathways, with a significance threshold of p < 0.05 [31]. Protein–protein interaction (PPI) networks were constructed via GeneMANIA (http://genemania.org), which integrates protein and genetic interactions, pathways, co-expression, colocalization, and domain similarity data to visualize functional associations among PGs. The cBioPortal platform was utilized to explore the relationship between PG mutations and patient prognosis, microsatellite instability (MSI), and tumor mutational burden (TMB) in BC.

Consensus cluster analysis

The intersection between these PFAS targets and the previously identified perineural invasion-associated genes (PGs) was defined as PFAS-related perineural invasion genes (PPGs) in BC, and visualized using the VennDetail package. Based on PPGs expression, consensus clustering was performed on TCGA-TNBC patients using the R package ConsensusClusterPlus with the following parameters: pItem = 0.8, clusterAlg = ‘km’, distance = “euclidean”, seed = 2024. The optimal number of clusters was determined by evaluating the cluster stability using cumulative distribution function (CDF) plots and cluster consensus histograms, along with corresponding heatmap visualizations. The k value with the highest clustering stability was selected as the final grouping scheme. Kaplan–Meier survival analysis was conducted using the survival and survminer packages to compare outcomes between TNBC subtypes. A 3D principal component analysis (PCA) plot was generated with the scatterplot3d package to visualize subgroup separation. Clinicopathological characteristics were also compared across the identified TNBC subtypes.

Construction and validation of the prognostic signature

Differential expression analysis between two TNBC clusters was performed using DESeq2, with |log₂FC|≥ 0.5 and adjusted p-value < 0.05 defining PFAS-mediated perineural invasion-related genes (PFASPNIs). We then integrated the RSF and StepCox algorithms to construct a PFAS and PNI-associated prognostic signature (PFASPNIsig). First, univariate Cox regression identified prognostic genes in the METABRIC-TNBC cohort. These genes were subsequently used to build prediction models via RSF and StepCox, which were then validated in the TCGA-TNBC cohort and the immunotherapy cohort IMvigor210. Model accuracy was assessed using ROC and decision curve analysis (DCA). A nomogram was developed and evaluated with calibration curves. The prognostic performance of PFASPNIsig was compared against multiple existing models across the METABRIC-TNBC, TCGA-TNBC, and IMvigor210 cohorts. All analyses were conducted within the “Mine1” package framework [32].

Drug sensitivity analysis

The Genomic of Drug Sensitivity in Cancer (GDSC) database (www.cancerRxgene.org) represents the largest publicly available resource linking cancer cell drug sensitivity to molecular markers of drug response [33]. Using the pRRophetic algorithm [34] applied to the TCGA-TNBC dataset, we estimated the half-maximal inhibitory concentration (IC50) to predict the sensitivity of different risk groups to clinically used BC therapeutics. The Cancer Immunome Atlas (TCIA, https://tcia.at/about) provides immunophenotype scores (IPS) for 20 cancer types, serving as a validated predictor of response to CTLA-4 and PD-1 blockade [35]. We obtained IPS data for TCGA-TNBC samples and compared scores among TNBC risk subgroups. Additionally, the Tumor Immune Dysfunction and Exclusion (TIDE) score (http://tide.dfci.harvard.edu/), which quantifies tumor immune evasion potential [36], was computed for each TCGA-TNBC sample to evaluate differences in TIDE immune scores across risk subgroups.

Mendelian randomization analysis

We performed two-sample Mendelian randomization (MR) to evaluate the causal effects of genetically predicted expression levels of the five PPGs (BIRC5, CCND1, CTSD, HNF4A, MMP9) on breast cancer (BC) risk. Single nucleotide polymorphisms (SNPs) associated with each of the five PPGs were selected from the eQTLGen Consortium (https://www.eqtlgen.org/), which provides cis-eQTL data from blood. The selection criteria were: (1) genome-wide significance threshold P < 1 × 10⁻5; (2) linkage disequilibrium clumping with r2 < 0.01 within a 10,000 kb window using the 1000 Genomes European reference panel; (3) exclusion of SNPs with minor allele frequency < 0.01. For each PPG, the number of independent SNPs retained as instrumental variables (IVs) was as follows: BIRC5: 2, CCND1: 8, CTSD: 19, HNF4A: 0, MMP9: 46. BC genome-wide association summary statistics were obtained from the IEU OpenGWAS database (GWAS ID: ieu-a—1126, based on the Breast Cancer Association Consortium). The outcome definition included all BC cases (both invasive and in situ) of European ancestry. The sample size consisted of 122,977 BC cases and 105,974 controls. The inverse variance weighted (IVW) method was used as the primary analysis to estimate causal effects. To assess the robustness of findings, we performed four sensitivity analyses: (1) MR-Egger to detect and adjust for directional pleiotropy; (2) weighted median estimator, which provides valid estimates if at least 50% of the weight comes from valid IVs; (3) simple mode and weighted mode approaches; (4) MR-PRESSO to identify and correct for outlier SNPs. Heterogeneity was assessed using Cochran’s Q statistic, and a leave-one-out analysis was conducted to evaluate whether any single SNP drove the causal estimate. All instrumental variables had F-statistics > 10, indicating no weak instrument bias. A two-sided P < 0.05 was considered statistically significant. All MR analyses were performed using the TwoSampleMR package (version 0.5.6) in R 4.3.2.

Molecular docking and molecular dynamics simulation

Molecular docking predicted interactions between PFAS (PFOA, PFOS) and CTSD. The human CTSD crystal structure (resolution < 2.0 Å) was obtained from UniProt, prepared by removing water and ligands in PyMOL, and docked using AutoDock Vina. The binding affinity (kcal/mol) was assessed from ten independent runs, with scores < –5 kcal/mol indicating stable binding. For the top complex (CTSD–Aurantio-obtusin), molecular dynamics simulations were performed in GROMACS using the CHARMM36 force field. The system was solvated, neutralized, energy-minimized, equilibrated, and subjected to a 100‑ns production run at 300 K and 1 bar.

Single-cell analysis

Seurat (v5.1.0) was used to analyze TNBC scRNA-seq data [37]. The raw UMI count matrix was filtered to remove genes detected in fewer than 5 cells and cells expressing fewer than 300 genes. High-quality cells were retained using thresholds of ≥ 500 UMIs, ≥ 100 genes, and < 20% mitochondrial gene content to preserve cellular heterogeneity for downstream analysis. Highly variable genes were identified from normalized data using the FindVariableFeatures function. Principal component analysis (PCA) was performed, followed by batch correction with Harmony and dimensionality reduction using UMAP. Cell clusters were identified using the FindClusters function with guidance from the ROGUE algorithm [38], and each cluster was manually annotated based on established marker genes. UMAP visualizations were generated using the scRNAtoolVis and scCustomize packages. Cluster-specific marker genes were identified with FindAllMarkers (min.pct = 0.25, logfc.threshold = 0.5). CellChat [39] was used to infer intercellular communication, leveraging its curated ligand-receptor interaction database and structural modeling. Pseudotime trajectory analysis was performed using Monocle2 [40]. Metabolic pathway activity scores for KEGG terms were computed with the scMetabolism package [41]. Spatial transcriptomic data (Visium_HD_FF_Human_Breast_Cancer_cloupe_008um) obtained from 10X Genomics (https://www.10xgenomics.com/datasets) were visualized using Loupe Browser 9.

Construction of GraphBAN graph neural network

To identify potential CTSD inhibitors from medicine-food homology (MFH) sources, we first systematically collected and curated a library of bioactive compounds. Based on the list of 106 MFH substances issued by the National Health Commission and State Administration for Market Regulation of China, we retrieved chemical ingredient data from the Traditional Chinese Medicine Systems Pharmacology Database (TCMSP), the BATMAN-TCM platform, and the Encyclopedia of Traditional Chinese Medicine (ETCM). Components meeting the criteria of oral bioavailability ≥ 30% and drug-likeness ≥ 0.18 were selected as active compounds. After removing duplicates, canonical SMILES notations were obtained from PubChem to construct a standardized MFH compound library for virtual screening. We then employed GraphBAN, an inductive graph neural network model [42], to predict interactions between these compounds and CTSD. The amino acid sequence of CTSD was retrieved from UniProt in FASTA format, while the compound SMILES served as input. Using a Python 3.8 environment with PyTorch 1.7.1 and DGL 0.7.1, GraphBAN extracted molecular features via GCN and ChemBERTa, and protein features via 1D-CNN and ESM, followed by fusion into a 128-dimensional representation. The model was trained using a teacher–student framework: the teacher module consisted of a graph autoencoder (learning rate 1e-3, 250 epochs), and the student module incorporated a bilinear attention network and conditional domain adversarial network (learning rate 1e-4, 50 epochs). Public datasets including BindingDB, BioSNAP, and KIBA were used for model training and optimization. Compounds with a predicted interaction probability > 0.5 were defined as key candidates. Finally, a "MFH source–active ingredient–target" network was constructed and visualized using Cytoscape.

Collection and organization of clinical samples from BC patients

The clinicopathological data and BC tissues used in this study were obtained from patients who underwent surgical treatment at the first people’s Hospital of Hefei. This study was approved by the Ethics Committee of the first people’s Hospital of Hefei [registration number: 2024—049—01]. A total of 173 patients were included in the study. Tumor tissues from 10 of these patients with TNBC were subsequently used for immunohistochemical staining (IHC) and immunofluorescence staining (IF).

Immunohistochemistry staining and immunofluorescence staining

Tissue samples from ten TNBC patients with or without PNI were collected for IHC and IFto examine the expression of Cathepsin D (Proteintech, 21,327—1-AP, 1:50) and CD68 (Proteintech, 25,747—1-AP, 1:50).

Cell lines and culture conditions

Human TNBC cells lines MDA-MB-231 (Procell, CL-0150) and BT549 (Procell, CL-0041) were maintained in basic RPMI 1640 medium (Procell, PM150110). Both culture media were supplemented with 10% fetal bovine serum (Procell, 164,210) and 1% penicillin–streptomycin (Beyotime, C0222). The cell cultures were incubated at 37°C in a humidified atmosphere containing 5% CO₂.

Small interfering RNA (siRNA) transfection

The siRNA sequences were designed based on a previous publication [28] and synthesized by Tsingke Biotechnology Co., Ltd. Two TNBC cell lines were cultured in 6-well plates until reaching 30–50% confluence. Transfection was performed by incubating 6 µL of Lipo6000 (Beyotime, C0526—0.5 ml), 50 nM of siRNA or negative control, and 250 µL of Opti-MEM for 20 min at room temperature. The resulting mixture was added dropwise to the cells, which were then incubated for 6 h. After replacing the medium with fresh complete medium, the cells were cultured for an additional 48 h before protein harvest and analysis.

Cell proliferation and colony formation assay

Cell proliferation was assessed using the Cell Counting Kit-8 (Beyotime, C0037). Briefly, cells were seeded in 96-well plates at a density of 4 × 103 cells per well and cultured in medium supplemented with 10% FBS at 37 °C. After 24 h, the cells were treated with various concentrations (100, 200, 400, 600, and 800 μM) of Aurantio-obtusin for 48 h. A 10% CCK-8 solution was then prepared in culture medium, added to each well, and incubated for 1 h at 37 °C. Absorbance was measured at 450 nm using a microplate reader. Clonogenic survival was evaluated by colony formation assay. Cells (3,000 per well) were seeded into 6-well plates and, after 24 h, continuously treated with 400 μM Aurantio-obtusin for 7 days. Subsequently, the cells were fixed with 10% formaldehyde for 30 min at room temperature and stained with 0.1% crystal violet for 2 h. Colonies consisting of more than 50 cells were counted.

Wound healing assay

TNBC cells were seeded in 6-well plates and cultured until reaching 90% confluence. A straight scratch was created in each well using a 10 µL pipette tip. After washing with PBS, the cells were cultured in serum-free medium. Following 48 h of incubation, the wounds were imaged under an inverted microscope.

Transwell migration assay

TNBC cells were seeded into the upper chamber of uncoated Transwell inserts (8 μm pore size; Corning Falcon, 3422) in 200 µL serum-free medium. The lower chamber was filled with 600 µL medium containing 10% FBS as a chemoattractant. After an appropriate incubation period, cells that migrated to the lower surface were fixed with 4% paraformaldehyde for 30 min and stained with crystal violet. Non-migrated cells in the upper chamber were gently removed. Images were acquired using an inverted microscope.

Western blot analysis

Proteins were extracted from cells using RIPA lysis buffer (Beyotime, P0013B). Protein concentration was determined with a BCA protein assay kit (Beyotime, P0009) and standardized accordingly. The protein extracts were mixed with SDS-PAGE loading buffer (5X) (Beyotime, P0015L) and boiled for 10 min. The proteins were separated by 10% SDS-PAGE and transferred onto a polyvinylidene fluoride (PVDF) membrane (Millipore, Temecula, CA, USA). The membrane was blocked with 5% skim milk for 2 h, followed by an overnight incubation at 4 °C with CTSD antibody (Proteintech, 21,327—1-AP, 1:1000). Subsequently, the membrane was incubated with HRP-conjugated Goat Anti-Rabbit IgG (H + L) (Proteintech, SA00001—2, 1:1000) for 2 h. Protein bands were visualized using a Chemidoc™ XRS + system with Image Lab™ software (Bio-Rad Laboratories, Inc., Hercules, CA, USA), and semiquantitative analysis was performed using the same software.

Statistical analysis

Statistical analyses were performed using R 4.3.2. Each experiment was repeated three times independently. Data were presented as mean ± SD unless otherwise specified. Quantitative data were compared by the Student’s t-test or one-way ANOVA, and Overall survival and lung metastasis-free survival analysis were conducted using the Kaplan–Meier method with the log-rank test. A p-value of less than 0.05 was considered statistically significant. In all cases, the p values are represented as follows: ∗ p < 0.05, ∗  ∗ p < 0.01, and ∗  ∗  ∗ p < 0.001.

Results

Identification and functional analysis of PGs in BRCA

Figure 1 illustrated the overall workflow of this study. We investigated the clinicopathological significance of PNIin BC and its associated molecular signatures. Analysis of 173 patients from our institutional cohort revealed that PNI was significantly correlated with larger tumor size (p < 0.05) and elevated Ki-67 index (Fig. 2A–C). To identify PNI-linked genes, we identified 8,249 differentially expressed genes (DEGs) between tumor and normal tissues from TCGA-BRCA dataset. Intersecting these with known PNI-related genes yielded 32 core PNI-associated DEGs (PGs) (Fig. 2D). Functional enrichment analysis indicated these PGs were involved in integrin-mediated adhesion, wound healing, and epithelial proliferation (Fig. 2E), with a protein–protein interaction network supporting their functional connectivity (Fig. 2F). Clinically, patients with PGs mutations showed a higher prevalence of the basal-like subtype (Fig. 2G), along with poorer disease-free and progression-free survival (Fig. 2H–I). The PGs-mutant group also exhibited increased microsatellite instability and tumor mutational burden (Fig. 2J–K). In summary, PNI in BC is characterized by distinct clinicopathological features and a specific gene expression profile linked to aggressive tumor behavior and poorer outcomes.

Fig. 1.

Overview of study design and analysis

Fig. 2.

Clinical Pathological Characteristics and Molecular Biomarkers of PNI in BRCA. A-C: The Age (A), diameter of tumor (B) and Ki-67 Score (C) in PNI positive and negative groups. (D). Venn diagram of intersection of DEGs with PNI related genes in TCGA-BRCA. (E). GO and KEGG enrichment analyses of 32 PGs in BRCA. (F). PPI network of 32 PGs. (G). Subtypes among TCGA-BRCA patients harboring or none- harboring 32 PGs mutations.. (H-I). The Kaplan–Meier curves of PFS and DFS among TCGA-BRCA patients harboring or none- harboring 32 PGs mutations. (J-K). MSI (J) and TMB (K) among TCGA-BRCA patients harboring or none- harboring 32 PGs mutations

PFAS exposure is associated with PNI in TNBC by driving an aggressive molecular subtype

This study aimed to elucidate the potential mechanisms through which perfluorinated alkyl substances (PFAS) contribute to PNIin TNBC, building on prior evidence linking PFAS exposure to metastatic progression. Since the 32 previously identified PNI-related genes (PGs) were predominantly associated with TNBC, we focused our investigation on this subtype. Using Venn analysis, we identified five PFAS-associated PNI-related genes (PPGs): BIRC5, CCND1, CTSD, HNF4A, and MMP9 (Fig. 3A), which are functionally implicated in apoptosis inhibition, cell cycle progression, proteolysis, transcriptional regulation, and extracellular matrix remodeling. To explore PFAS-related PNI heterogeneity in TNBC, we performed unsupervised clustering based on these five PPGs, which stratified tumors into two distinct molecular subgroups: Subtype 1 (n = 44) and Subtype 2 (n = 55) (Supplementary Fig. 1A, B). Principal component analysis confirmed clear transcriptomic separation /. between the two subtypes (Supplementary Fig. 1C). Survival analysis revealed that patients in Subtype 1 had significantly shorter overall survival than those in Subtype 2 (Fig. 3B). Moreover, Subtype 1 was significantly associated with advanced N stage and poorer survival status (Supplementary Fig. 1D), indicating more aggressive clinical behavior. We next identified 209 downregulated and 307 upregulated genes (adjusted p < 0.05, |log2FC|≥ 1) between Subtype 1 and Subtype 2. Gene Set Enrichment Analysis (GSEA) showed significant enrichment of pathways related to aggressive phenotypes, including BC metastasis, non-integrin ECM interactions, focal adhesion, TGFβ-induced EMT, and ECM receptor interaction (Fig. 3C). In addition, Subtype 1 showed elevated proliferative potential, as evidenced by GSVA enrichment of E2F targets, G2M checkpoint, and MYC targets (Fig. 3D). Further profiling via ssGSEA indicated that Subtype 1 exhibited stronger activity in EMT pathways (Fig. 3E). Furthermore, the expression of both CTSD and CCND1 showed positive correlations with the EMT score (Fig. 3F–G). Collectively, these results indicate that Subtype 1 represents a clinically aggressive TNBC subgroup with poor prognosis, characterized by a distinct PFAS-associated PNI gene expression signature, enhanced proliferative drive, and EMT activation.

Fig. 3.

Molecular and Clinical Characterization of Two Novel PPGs-Based Subtypes in TNBC. (A). Identification of 5 PPGs through the intersection of PGs and PFAS-related genes. (B). Kaplan–Meier curves for overall survival of patients stratified into two PPGs subtypes. (C). Gene Set Enrichment Analysis highlighting the top signaling pathways that are differentially activated between the two subtypes. (D). GSVA enrichment scores for the Hallmark gene sets, uncovering subtype-specific metabolic and oncogenic pathways. (E). Evaluation of EMT phenotypes using ssGSEA with three established EMT gene signatures in each subtype. (F-G). Correlation analysis of the composite EMT score with the expression of key genes (F) CTSD and (G) CCND1 in the TCGA-TNBC dataset. The solid line represents the linear regression fit, with the Pearson correlation coefficient (R. and p-value indicated

Construction and validation of a prognostic model and predictive nomogram based on 5 PPGs-associated subtypes

To investigate the potential role of 5 PPGs in the progression and clinical outcomes of TNBC, we developed a prognostic model using differentially expressed genes (DEGs) derived from two subtypes previously classified based on these PPGs. The model was constructed using data from the METABRIC cohort, which included 166 TNBC patients with complete overall survival (OS) information. External validation was performed using 97 TNBC patients from the TCGA-BRCA cohort and 326 patients from the IMvigor210 cohort who had received immunotherapy. Through a combination of random survival forest analysis and Cox proportional hazards regression, we identified eight genes significantly associated with prognosis (Fig. 4A). Using the median risk score as the cutoff, patients in the training set were stratified into high-risk and low-risk groups. Kaplan–Meier survival analysis revealed that patients in the high-risk group had significantly shorter overall survival compared to those in the low-risk group (Fig. 4B). The prognostic model demonstrated strong predictive accuracy, with time-dependent receiver operating characteristic (ROC) analysis showing area under the curve (AUC) values of 0.72, 0.74, and 0.80 for 1‑, 3‑, and 5‑year overall survival, respectively (Fig. 4D–F). Consistent results were observed in the TCGA-BRCA validation cohort (Fig. 4C–F). A meta-analysis further confirmed the significant association between the risk score and poor prognosis in TNBC patients (Fig. 4G). Our 8-gene signature demonstrated superior predictive accuracy compared to other existing models in both the internal training and external validation cohorts (Supplementary Fig. 2). Subsequently, We constructed a nomogram incorporating clinical parameters and the risk score to predict 1‑, 3‑, and 5‑year overall survival in TNBC patients from the METABRIC cohort (Fig. 5A). Calibration plots indicated close agreement between the nomogram-predicted probabilities and actual observed outcomes (Fig. 5B). Decision curve analysis (DCA) further demonstrated the favorable clinical utility of the nomogram across a range of threshold probabilities (Fig. 5C). In summary, our study establishes and validates a 5 PPGs-derived gene signature that effectively stratifies TNBC patients into distinct risk subgroups, with significant implications for prognostic assessment and treatment decision-making.

Fig. 4.

Development and Multi-Cohort Validation of a PPGs-Derived Prognostic Signature. (A). The top panel shows the scatter plot of patients in the METABRIC-TNBC cohort, categorized by risk score and annotated with their vital status. The bottom panel presents the expression heatmap of the final 8 prognostic genes selected by the RSF and StepCox algorithm. (B-C). Kaplan–Meier curves analyzing OS between the high- and low-risk groups in the (B) METABRIC-TNBC and (C) TCGA-TNBC cohorts. (D-F). Time-dependent receiver operating characteristic (ROC) curves evaluating the predictive accuracy of the risk model for (D) 1-year, (E) 3-year, and (F) 5-year OS in the METABRIC-TNBC and TCGA-TNBC cohorts. (G). Forest plots from meta-analysis and hazard ratios from univariate Cox regression consistently validate the risk score as a significant prognostic factor across all three independent cohorts

Fig. 5.

A Clinical Nomogram for Survival Prediction and the Association of the Risk Model with therapy Efficacy. A. (A) nomogram integrating the risk score, tumor grade, and disease stage to predict 1-, 5-, and 8-year overall survival for patients in the METABRIC-TNBC cohort. (B). Calibration curves for the nomogram. The 45-degree dotted line represents the ideal prediction, and the solid line represents the actual performance of the nomogram. (C). DCA evaluating the clinical net benefit of the risk model for predicting 1-, 5-, and 8-year OS. (D). Kaplan–Meier curves comparing OS between the high- and low-risk groups in the IMvigor210 cohort. E–F. Time-dependent ROC curves assessing the predictive accuracy of the risk score for (E) 1-year and (F) 3-year OS in the IMvigor210 cohort. (G). TIDE scores between the high- and low-risk groups. (H). Scores for the “Exclusion”, "IFN-gamma" (IFNg), and “Merck18” signatures in the two risk groups. (I). Violin plots of the IPS across risk groups. (J). Box plots comparing the estimated IC50 of commonly used drugs for BC patients between the two risk groups

High-risk group exhibits immunosuppressive phenotype and therapy resistance

Recent evidence suggests that perineural invasion may influence tumor responsiveness to immunotherapy. To further investigate this, we utilized the IMvigor210 cohort—a dataset of urothelial carcinoma (BLCA) patients treated with immune checkpoint inhibitors—to characterize the immunotherapeutic features associated with high- and low-risk groups stratified based on a novel prognostic signature. Survival analysis revealed that patients in the high-risk group had significantly shorter overall survival compared to those in the low-risk group (Fig. 5D). The prognostic model demonstrated high sensitivity and specificity, with area under the receiver operating characteristic (ROC) curve values of 0.603 and 0.615 for predicting 1- and 3-year survival, respectively (Fig. 5E–F). Further exploration indicated that high-risk patients exhibited significantly higher TIDE scores (Fig. 5G), suggesting enhanced potential for immune evasion. Consistently, this group showed elevated exclusion scores but lower IFNG and Merck18 gene expression scores (Fig. 5H), implying impaired T-cell function and interferon-γ response. Additionally, lower IPS scores were observed in high-risk patients for both CTLA4( −)PD1( +) and CTLA4( +)PD1( +) configurations (Fig. 5I), indicating reduced responsiveness to immune checkpoint blockade. We also assessed the sensitivity of each risk group to conventional chemotherapeutic agents commonly used in BRCA. The high-risk group exhibited significantly reduced sensitivity to Cisplatin, Oxaliplatin, Paclitaxel, Cytarabine, Docetaxel, and Vinblastine (all p < 0.05; Fig. 5J). In summary, our results indicate that the high-risk group defined by our model is associated with an immunosuppressive phenotype and reduced survival, along with increased resistance to both immunotherapy and conventional chemotherapy.

Genetic and structural evidence identifies CTSD as a causal mediator in PFAS-driven breast cancer metastasis

For the five PPGs, we initially identified a total of 75 SNPs as candidate instruments. After LD clumping and excluding SNPs directly associated with BC (P < 5 × 10⁻⁸), we retained 7 SNPs across the five genes. The number of independent SNPs retained as instrumental variables (IVs) for each gene was: BIRC5: 1, CCND1: 2, CTSD: 2, HNF4A: 0 (no valid instrument), MMP9: 2. All F-statistics exceeded 10, confirming instrument strength. Using the IVW method, we identified two genes with nominally significant associations with BC risk (Fig. 6A). Genetically predicted higher CCND1 expression was associated with a reduced BC risk (OR = 0.8915, 95% CI: 0.8113–0.9797, P = 0.017), suggesting a potential protective role. In contrast, genetically predicted higher CTSD expression was associated with an increased BC risk (OR = 1.0944, 95% CI: 1.0298–1.1629, P = 0.004), indicating a potential risk-promoting effect. For the remaining three genes (BIRC5, HNF4A, MMP9), no significant associations were observed (P > 0.05). Sensitivity analyses showed no evidence of horizontal pleiotropy (MR-Egger intercept P > 0.05 for all) and no significant heterogeneity (Cochran’s Q P > 0.05). Leave-one-out analysis confirmed that no single SNP drove the observed associations (Supplementary Fig. 3C). MR-PRESSO did not identify any outlier SNPs. Scatter plots illustrated a positive correlation between weighted genetic risk scores and BC incidence (Supplementary Fig. 3A), and forest plots reinforced consistent directional effects across instruments (Supplementary Fig. 3D). Funnel plots revealed no significant asymmetry, suggesting minimal heterogeneity and low likelihood of directional pleiotropy (Supplementary Fig. 3B). Together, these findings support a potential causal association between CTSD expression and BC risk (Tables S2–S4). Molecular docking revealed strong binding affinities between CTSD and the PFAS compounds PFOA and PFOS, with calculated binding energies below –8 kcal/mol (Fig. 6B–C). Compared with normal breast tissues, CTSD expression was markedly elevated in breast carcinoma samples (Fig. 6D). Furthermore, among tumor tissues, CTSD expression was higher in TNBC with PNI than in those without (Fig. 6E). In TNBC cell lines, knockdown of CTSD significantly suppressed cell migration (Fig. 6F–H).

Fig. 6.

Causal Inference and Molecular Interactions of CTSD in BC. (A). Forest plot showing NR estimates of the causal effects of 5 key PPGs on BC risk. (B). Predicted binding modes between CTSD and (left) PFOA or (right) PFOS, as revealed by molecular docking. (C) Binding affinities (kcal/mol) of PFOA and PFOS to CTSD, presented in a lollipop chart. (D). IHC from the HPA database comparing CTSD expression in normal breast tissue and BC. (E). Representative IHC images of CTSD and IF images of CTSD and CD68 in TNBC with or without neural invasion. (F). Western blot analysis of CTSD expression in MDA-MB-231 and BT-549 cells after CTSD knockdown. (G-H). Transwell migration assay (G) and wound healing assay (H) in MDA-MB-231 and BT-549 cells after CTSD knockdown. (**p < 0.01, ***p < 0.001)

CTSD enrichment in myeloid cells associates with poor prognosis markers and treatment resistance

To characterize the cellular expression profile of CTSD in TNBC, we analyzed single-cell RNA sequencing data from eight TNBC samples (GSE161529). After quality control, 55,917 high-quality cells were retained (Supplementary Fig. 4A). Unsupervised clustering and UMAP visualization using the Seurat workflow identified eight major cell types (Fig. 7A), including T cells (CD2, CD3D, TRAC, TRBC2), B cells (CD79A, CD79B, MS4A1), epithelial cells (EPCAM, CDH1, KRT8, KRT18), fibroblasts (DCN, COL1A1, COL1A2), pericytes (RGS5, ACTA2, TAGLN), myeloid cells (CD14, FCN1, C1QC), plasma cells (JCHAIN, MZB1, IGHG1), and endothelial cells (PECAM1, CDH5, VWF, CLDN5) (Supplementary Fig. 4B–D). Substantial inter-sample heterogeneity in cellular composition was observed (Supplementary Fig. 4E–F). CTSD expression was predominantly enriched in myeloid cells (Fig. 7B). Re-clustering of 6,082 myeloid cells revealed multiple subpopulations (Fig. 7C). Based on CTSD expression (Fig. 7D–E), clusters 0, 4, 5, and 6 were classified as CTSD-positive myeloid (CTSD_pMye), while the remaining clusters were designated CTSD-negative myeloid (CTSD_nMye) (Fig. 7F). The CTSD_pMye subset exhibited elevated expression of myeloid markers linked to poor prognosis, including SPP1, CD86, CD68, and CCL2 (Fig. 7G). This finding was validated in an independent dataset (GSE266919), where myeloid cells were partitioned into 14 subclusters (Fig. 7H–I). CTSD_pMye (clusters 0, 1, 3, 5, 9) and CTSD_nMye were similarly defined by CTSD expression (Fig. 7J–K), and CTSD_pMye again showed upregulated expression of pro-tumorigenic genes such as SPP1, MSR1, CD68, and CCL2 (Fig. 7L). Clinically, CTSD_pMye abundance was associated with treatment response: CTSD expression was significantly higher in patients with stable or progressive disease than in those with partial response (Fig. 7M–N).

Fig. 7.

Characterization of CTSD-Enriched Myeloid Subsets and Their Association with Therapy Response in scRNA-seq Datasets. (A). UMAP visualization of the annotated cell clusters in the GSE161529 scRNA-seq dataset. (B). UMAP showing the expression distribution of CTSD across all cell types in GSE161529. (C). UMAP plot of re-clustered myeloid cells extracted from the GSE161529 dataset. (D). Expression distribution of CTSD across the identified myeloid cell subpopulations in GSE161529. (E). Dot plot illustrating the expression level and proportion of CTSD in each myeloid subcluster. (F). UMAP of myeloid cells from GSE161529, categorized into CTSD-Positive (clusters 0, 4, 5, 6) and CTSD-Negative groups based on expression levels. (G). Dot plot comparing the expression of macrophage-related genes (CCL2, CD68, CD86, SPP1) and CTSD between the CTSD-Positive and CTSD-Negative groups. (H). UMAP of myeloid cells from the GSE266919 scRNA-seq dataset after reduction and clustering. (I). UMAP plot showing the different clusters of myeloid cells in the GSE266919 dataset. (J). Expression distribution of CTSD across myeloid cell in GSE266919. (K). UMAP of GSE266919 myeloid cells, stratified into CTSD-Positive (clusters 0, 1, 3, 5, 9) and CTSD-Negative groups. (L). Dot plot displaying the expression of key genes (CCL2, CD68, MSR1, SPP1, CTSD) in the CTSD-Positive vs. CTSD-Negative groups. (M). Dot plot showing the expression of the gene panel across myeloid cells from patients with progressive disease (PD), partial response (PR), and stable disease (SD). (N). Dot plot comparing the expression of the gene panel between responders and non-responders to therapy

Differentiation into CTSD-positive myeloid cells correlates with poor prognosis and malignant progression

We further applied pseudotime trajectory analysis to examine the relationship between CTSD expression and differentiation states in TNBC myeloid cells. The analysis revealed that CTSD_pMye were predominantly located at the terminal end of the differentiation trajectory compared to CTSD_nMye, indicating a more mature myeloid differentiation stage (Fig. 8A–B). To validate the spatial distribution of this subset, we analyzed high-resolution spatial transcriptomic data from a BC specimen. Macrophages were identified based on co-expression of LYZ and CD68, and further classified into CTSD_Positive_macrophage and CTSD_Negative_macrophage using a log2-transformed CTSD expression threshold of ≥ 1.5 (Fig. 8C–F). CTSD_Positive_macrophage exhibited significantly higher expression of CD68, SPP1, and CCL2 than the CTSD_Negative group (Fig. 8G). To assess the clinical relevance of myeloid subpopulations, we evaluated their infiltration levels in multiple TNBC cohorts using ssGSEA based on subpopulation-specific gene signatures. High infiltration of clusters 3 and 8 from the CTSD_nMye group was associated with favorable patient prognosis (Fig. 8H), implying a protective role for CTSD-low myeloid cells. Pseudotime dynamics analysis showed that both CTSD and CD68 expression increased significantly along the differentiation trajectory (Fig. 8I–J). Immunofluorescence staining further confirmed enhanced co-localization of CTSD and CD68 in myeloid cells within perineural invasion regions (Fig. 6E). Metabolic profiling indicated that CTSD_pMye displayed elevated activity in multiple pathways, including arginine biosynthesis, fatty acid biosynthesis, and arginine, proline, alanine, aspartate, and glutamate metabolism (Fig. 8K–N).

Fig. 8.

CTSD-Positive Myeloid Cells Represent a Terminally Differentiated State Linked to Metabolic Reprogramming and Poor Survival in TNBC. (A). Trajectory inference depicting the developmental path of myeloid cells in GSE161529. (B). The CTSD-Positive cells are positioned at the terminal end of the pseudotime trajectory. (C-F). Spatial transcriptomic analysis of a BC sample. (C) H&E staining; (D) Macrophages were identified by co-expression of LYZ and CD68; Spatial distribution of CTSD-Positive (E) and CTSD-Negative (F) macrophages, classified by a log2-normalized expression cutoff of 1.5. (G). Violin plots confirming higher expression of CD68, SPP1, and CCL2 in CTSD-Positive macrophages. (H). Survival analysis of clusters 3 and 8 (characteristic of the CTSD-Negative phenotype) in METABRIC-TNBC cohort. (I-J). Gene expression dynamics along pseudotime show significant positive correlations for (I) CTSD and (J) CD68. (K-N). Metabolite pathway analysis reveals significant enrichment of (K) Arginine biosynthesis, (L) Fatty acid biosynthesis, (M) Arginine and proline metabolism, and (N) Alanine, aspartate and glutamate metabolism pathways in CTSD-Positive versus CTSD-Negative myeloid cells

Identification of CTSD-positive malignant epithelial cell subpopulations and their role in promoting metastasis and immunosuppressive microenvironments

From eight TNBC samples in the GSE161529 dataset, we isolated 26,667 high-quality epithelial cells through clustering and CopyKat-based malignancy inference (Fig. 9A). Malignant epithelial cells were subdivided into 11 subclusters (Fig. 9B) and classified into CTSD‑positive (CTSD_pEpi) and CTSD‑negative (CTSD_nEpi) groups based on CTSD expression (Fig. 9C–D). Cluster 4 exhibited the highest CTSD enrichment and was linked to poor TNBC prognosis (Fig. 9E). ssGSEA using these genes showed positive correlation with both epithelial–mesenchymal transition (EMT) and T‑cell exclusion scores (Fig. 9F–G). Functional analysis of its marker genes revealed enrichment in cell adhesion, focal adhesion, and endoplasmic reticulum stress pathways (Fig. 9H). Cell–cell communication analysis revealed stronger interactions between CTSD_pEpi and cancer‑associated fibroblasts (CAFs) (Fig. 9I–J). Specifically, fibroblasts communicated with CTSD_pEpi via the FGF pathway, but not with CTSD_nEpi (Fig. 9K–L). Although both epithelial subtypes interacted with myeloid cells through MIF signaling, CTSD_pEpi showed stronger communication (Fig. 9M–N). Metabolically, CTSD_pEpi exhibited elevated activity in glycolysis/gluconeogenesis, fatty acid biosynthesis, and alanine, aspartate, and glutamate metabolism compared to CTSD_nEpi (Fig. 9O–Q).

Fig. 9.

Identification of a Pro-Metastatic and Immunosuppressive CTSD + Malignant Epithelial Subpopulation Driven by Metabolic Rewiring and Unique Microenvironment Crosstalk. (A). UMAP visualization of 26,667 high-quality epithelial cells in the GSE161529 dataset, annotated by CopyKat-inferred malignancy. (B). Re-clustering of malignant epithelial cells. (C). Malignant epithelial cells were stratified into CTSD-Positive and CTSD-Negative groups based on expression levels. (D). CTSD is predominantly enriched in cluster 4. (E). The cluster 4 signature score is associated with poor patient prognosis. (F-G). The cluster 4 signature score correlates positively with (F) EMT score and (G) T-cell exclusion score, indicating roles in metastasis and immunosuppression. (H). GO enrichment analysis reveals that marker genes of cluster 4 are significantly enriched in pathways related to cell adhesion, focal adhesion, and endoplasmic reticulum stress. (I). Cell–cell communication network diagram. (J). Compared to CTSD-Negative cells, CTSD-Positive malignant epithelial cells exhibit significantly stronger interactions with cancer-associated fibroblasts (CAFs). (K-L). Cell communication between CTSD-Positive cells and CTSD-Positive cells and Fibroblasts in the FGF signaling pathway. (M-N). Cell communication between CTSD-Positive cells and CTSD-Positive cells and myeloid cells in the MIF signaling pathway. (O-Q). Metabolic analysis shows enhanced activity in (O) Glycolysis / Gluconeogenesis, (P) Fatty acid biosynthesis, and (Q) Alanine, aspartate and glutamate metabolism in CTSD-Positive versus CTSD-Negative malignant cells

Identification of Aurantio-obtusin as a CTSD inhibitor in TNBC

Based on the pivotal role of CTSD in TNBC perineural invasion, we screened for potential CTSD inhibitors. Using a GraphBAN graph neural network, we evaluated interactions between 106 bioactive ingredients from medicine‑food homology (MFH) sources and CTSD. Compounds with a CPI > 0.5 in at least two of the BindingDB, BioSNAP, and KIBA databases were defined as high‑confidence candidates, resulting in 33 candidates (Fig. 10A; Table S5). A regulatory “MFH source–ingredient–CTSD” network was constructed with Cytoscape (Fig. 10B). Among them, Aurantio‑obtusin, showing CPI > 0.8 in both BioSNAP and KIBA, was identified as a core candidate. Molecular docking revealed a binding free energy of –7.5 kcal/mol between Aurantio‑obtusin and CTSD, indicating strong binding potential (Fig. 10C). To assess dynamic binding stability, we performed 100 ns molecular dynamics simulations. The RMSD of the CTSD–Aurantio‑obtusin complex and free CTSD stabilized after 50 ns at ~ 1.5 nm (Fig. 10D–E). The complex exhibited low RMSF (Fig. 10F), indicating minimal conformational fluctuation. The radius of gyration decreased from ~ 4 nm to ~ 3 nm (Fig. 10G), reflecting enhanced compactness. The solvent‑accessible surface area decreased from 335 nm2 to 315 nm2 and equilibrated at 60 ns (Fig. 10H), suggesting a more buried binding interface. Hydrogen bond analysis confirmed stable bond number and strength during simulation (Fig. 10I). The free energy landscape indicated a transition from an unfolded to a stably folded conformation (Fig. 10J). We next evaluated the anti‑tumor activity of Aurantio‑obtusin in TNBC cells. It inhibited cell proliferation (Fig. 10K) and CTSD protein expression (Fig. 10L) in a concentration‑dependent manner, and suppressed colony formation (Fig. 10M) and cell migration in Transwell and wound healing assays (Fig. 10N–O), effects potentially linked to CTSD downregulation.

Fig. 10.

Computational screening of food-derived active ingredients targeting CTSD using GraphBAN and analysis of their interactions. (A). Heatmap of the compound–protein interaction (CPI) probabilities between 33 key ingredients and CTSD predicted by GraphBAN across the BindingDB, BioSNAP, and KIBA datasets. (B). Regulatory network of " MFH –ingredients – CTSD". C. Molecular docking pose of CTSD with the core candidate compound Aurantio-obtusin, with the calculated binding free energy indicated. (D-E). Root mean square deviation (RMSD) analysis of the CTSD-Aurantio-obtusin complex (D) and the CTSD protein alone (E) over 100 ns of molecular dynamics simulation. (F). Root mean square fluctuation (RMSF) of the CTSD-Aurantio-obtusin complex, reflecting conformational flexibility of protein residues during the simulation. (G). Radius of gyration (Rg) of the CTSD-Aurantio-obtusin complex, indicating the overall compactness of the complex structure. (H). Solvent-accessible surface area (SASA) of the CTSD-Aurantio-obtusin complex, characterizing changes in surface exposure during complex formation. (I). Number of hydrogen bonds formed between CTSD and Aurantio-obtusin during the simulation, highlighting the stability of the protein–ligand interaction. (J). Free energy landscape of the CTSD-Aurantio-obtusin complex, depicting the conformational space and energy profile. (K–O). Effects of DMSO and different concentrations of Aurantio-obtusin on MDA-MB-231 and BT-549 cells were assessed by CCK-8 assay (K), Western blot (L), colony formation assay (M), wound healing assay (N), and Transwell assay (O). (**p < 0.01, ***p < 0.001)

Discussion

This study integrates multi-omics data to suggest a novel pathway in which the environmental pollutant PFAS is associated with neural invasion and malignant progression in TNBC, potentially involving disruption of CTSD function. We have not only established CTSD as a key causal molecule linking PFAS exposure to high-risk TNBC, but also demonstrated at the single-cell level that CTSD is enriched in specific myeloid cell and malignant epithelial cell subsets. By remodeling the immunosuppressive and metabolic microenvironment, these cells collectively promote the tumor’s perineural invasion phenotype and therapy resistance. Additionally, we have identified Aurantio-obtusin as a potential targeting strategy for CTSD. These findings provide new insights into the molecular pathology of environmental carcinogenesis and establish CTSD as a highly promising therapeutic target.

Previous epidemiological studies have linked PFAS exposure to several cancers, such as those of the liver and thyroid. These reports suggest that PFAS may enhance tumor metastatic potential [12–15]. In this study, we connect PFAS exposure to a specific molecular subtype of TNBC—Subtype 1—which shows high perineural invasion (PNI) potential. This subtype is defined by five PFAS-associated PNI genes: BIRC5, CCND1, CTSD, HNF4A, and MMP9. It exhibits strong EMT activity, high proliferation, and poor prognosis. CCND1 not only is associated with uncontrolled cell proliferation—essential for tumor growth and local invasion—but also plays non-canonical roles in cell motility and invasion. In TNBC, the EMT driver FOXQ1 and its binding partner RAPH1‑i3 upregulate CCND1 through STAT3 activation. This leads to radio-resistance, indicating CCND1’s role in an EMT–radio-resistance axis [43]. These findings align with the high EMT score seen in Subtype 1. We also found that both CTSD and CCND1 expression positively correlates with EMT scores. This supports their involvement in early metastatic invasion. Other genes, including BIRC5, HNF4A, and MMP9, have also been linked to malignant progression through various pathways [44–46]. Together, these results reveal how PFAS exposure is associated with an aggressive TNBC phenotype at the molecular level. They provide a biological basis for earlier epidemiological observations.

Among the five key genes, we employed MR analysis and identified CTSD as a molecule potentially linking PFAS exposure to TNBC risk, providing genetic evidence supporting an association. CTSD, a widely expressed acidic lysosomal protease [47], is critically involved in various cancers and BC metastasis, and its expression—directly induced by estrogen and growth factors—significantly enhances metastatic potential [48]. Mechanistically, CTSD is associated with tumor cell migration, invasion, and metastasis by upregulating intercellular adhesion molecule-1 [49]. Clinical data further support its pro-tumorigenic role: positive CTSD expression is significantly correlated with high histological grade, increased risk of recurrence, distant metastasis, and poor prognosis in BC [50]. Additionally, in neuroblastoma, CTSD interacts with the EGFR signaling pathway, and low CTSD levels are associated with aggressive behavior and unfavorable outcomes, suggesting its involvement in cell cycle and proliferation regulation during neural invasion [51]. Together, this evidence indicates that CTSD not only acts as a key driver of BC metastasis but may also indirectly facilitate perineural invasion by modulating progression in nerve-related tumors.

Molecular docking revealed stable binding of PFAS (e.g., PFOA, PFOS) to CTSD, suggesting PFAS may directly perturb CTSD function. As environmental pollutants, PFAS promote breast cancer cell migration [16]. Other contaminants like nanoplastics and arsenic disrupt cellular homeostasis by inhibiting CTSD, leading to blocked autophagy or impaired immune surveillance [25–27]. Thus, in TNBC, PFAS may similarly disrupt tumor cell homeostasis and anti-tumor immunity, facilitating perineural invasion (PNI) and metastasis.

Single-cell analysis identified specific enrichment of CTSD in a myeloid subpopulation within the TNBC microenvironment. These CTSD⁺ myeloid cells express immunosuppressive markers (CD68, SPP1, CCL2) and expand in therapy-resistant patients. Notably, CD68 and CTSD co-localize in TNBC tissues with PNI, indicating CTSD⁺ myeloid cells are a specialized macrophage subset. CCL2 from tumor-associated macrophages is associated with metastasis via Akt/β-catenin [52], while SPP1⁺ macrophages drive tumor progression and immunosuppression [53], with poor prognosis in BC [54] and linked to chemotherapy failure in TNBC [55]. Therefore, CTSD⁺ myeloid cells are a key driver of immunosuppression and therapy resistance. Prior studies show CTSD-targeting antibodies polarize macrophages to an M1 phenotype, sensitizing TNBC to immunotherapy [24]. Our work reinforces CTSD’s therapeutic potential and elucidates how its inhibition may enhance immunotherapy efficacy.

CTSD⁺ myeloid cells exhibit a unique immunosuppressive metabolic profile, characterized by highly active fatty acid metabolism—similar to lipid-associated macrophages (LAMs) that impair anti-PD-1 efficacy [56]. This metabolic rewiring supports both their survival and M2-like polarization [57, 58]. These cells also show enhanced glutamine and arginine metabolism. Glutamine metabolism may further reinforce immunosuppression, as its targeting can repolarize macrophages toward an anti-tumor M1 phenotype [59]. Arginine metabolism depletes local arginine, impairing T cell cytotoxicity and promoting tumor progression [60, 61]. Together, these metabolic features enable CTSD⁺ myeloid cells to sustain themselves while driving immunosuppression and therapy resistance.

CTSD⁺ epithelial cells form a distinct subpopulation linked to poor prognosis, EMT, and an immune-excluded phenotype. These cells actively communicate with the tumor microenvironment, particularly via FGF signaling with CAFs, creating a bidirectional loop: CTSD⁺ cells may activate CAFs, which in turn promote tumor progression via FGF [62]. CTSD itself drives proliferation, fibroblast growth, angiogenesis, and anti-apoptosis [63]. Enhanced MIF signaling also strengthens interactions between CTSD⁺ epithelial cells and myeloid cells, reinforcing immunosuppression. Additionally, this subpopulation exhibits heightened glycolysis, fatty acid, and amino acid metabolism, supplying energy and precursors to support proliferation, signaling crosstalk, and therapy resistance.

Based on the central role of CTSD, we further screened potential natural inhibitors from medicinal-edible substances. Computational modeling and molecular dynamics simulations revealed that Aurantio-obtusin can form a stable complex with CTSD. Preliminary cellular experiments confirmed that Aurantio-obtusin effectively suppresses the proliferation and migration of TNBC cells, accompanied by downregulation of CTSD protein expression. These findings identify Aurantio-obtusin as a promising lead compound for CTSD-targeted intervention strategies.

While this study systematically reveals a potential pathway through which PFAS exposure is associated with perineural invasion and malignant progression in TNBC by perturbing CTSD—using integrated multi-omics, Mendelian randomization, and single-cell analyses—several limitations remain. First, the specific molecular steps by which PFAS directly regulates CTSD expression or activity require further biochemical validation. Second, our study lacks direct PNI annotation in the public datasets. The PNI gene signatures, derived mainly from gastric cancer, may not fully capture TNBC-specific PNI biology. Thus, our analyses might reflect general tumor aggressiveness rather than PNI-specific mechanisms. To mitigate this, we focused on functionally validated PNI genes and validated key findings in our own pathologically annotated cohort. Future prospective TNBC cohorts with detailed PNI assessment are needed to confirm specificity. Third, TNBC definition was inconsistent across datasets: TCGA-BRCA used clinical receptor status (ER − /PR − /HER2 −), while METABRIC used PAM50 basal-like subtype as a surrogate. These definitions are correlated but not identical—some receptor-negative tumors are not basal-like, and some basal-like tumors express low hormone receptors. This heterogeneity may introduce variability. We have specified the definition for each dataset in Methods and acknowledge this limitation. Additionally, our experimental validation is limited to in vitro assays demonstrating that CTSD knockdown reduces TNBC cell migration. We did not directly test whether PFAS exposure upregulates CTSD or induces PNI‑related phenotypes in co‑culture or in vivo models. Therefore, while our findings support a functional link between CTSD and TNBC cell motility, direct evidence that PFAS drives PNI through CTSD is lacking. Future studies using PFAS‑treated TNBC cells in nerve co‑culture systems or animal models are needed to establish this causal relationship.

Conclusion

This study integrates multi-omics data to suggest a novel pathway in which the environmental pollutant PFAS is associated with neural invasion and malignant progression in TNBC, potentially involving disruption of CTSD function. We identified CTSD as the central hub connecting PFAS exposure to TNBC risk, acting through multiple pro-tumor mechanisms within the tumor microenvironment. Single-cell analysis uncovered two key CTSD-enriched subpopulations: metabolically active, immunosuppressive CTSD⁺ myeloid cells and CTSD⁺ malignant epithelial cells showing enhanced crosstalk with cancer-associated fibroblasts and myeloid cells. However, direct experimental evidence linking PFAS exposure to CTSD‑mediated PNI remains to be established. These populations cooperate through elevated SPP1/CCL2 expression and activated fatty acid/glutamine metabolism, collectively fostering a microenvironment conducive to tumor proliferation, EMT, immune evasion and therapy resistance. Our findings establish CTSD as a promising prognostic biomarker and therapeutic target. Targeting CTSD may simultaneously inhibit tumor malignancy and remodel the immunosuppressive microenvironment, providing new strategies against TNBC progression.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

PFAS

Per- and polyfluoroalkyl substances

PFOA

Perfluorooctanoic acid

PFOS

Perfluorooctane sulfonate

BC

Breast cancer

TNBC

Triple-negative breast cancer

PNI

Perineural invasion

CTSD

Cathepsin D

GEO

Gene Expression Omnibus

GSE

GEO Series accession number

ER

Estrogen receptor

PR

Progesterone receptor

HER2

Human epidermal growth factor receptor 2

ICBs

Immune checkpoint blockers

TME

Tumor microenvironment

EMT

Epithelial-mesenchymal transition

DEGs

Differentially expressed genes

GO

Gene ontology

KEGG

Kyoto Encyclopedia of Genes and Genomes

PPI

Protein-protein interaction

MSI

Microsatellite instability

TMB

Tumor mutational burden

PCA

Principal component analysis

RSF

Random survival forest

DCA

Decision curve analysis

ROC

Receiver operating characteristic

AUC

Area under the curve

OS

Overall survival

GDSC

Genomics of drug sensitivity in cancer

TCIA

The Cancer Immunome Atlas

TIDE

Tumor Immune Dysfunction and Exclusion

MR

Mendelian randomization

IVW

Inverse variance weighted

SNPs

Single nucleotide polymorphisms

scRNA-seq

Single-cell RNA sequencing

UMAP

Uniform manifold approximation and projection

ROGUE

Robust gene expression analysis

GSEA

Gene set enrichment analysis

GSVA

Gene set variation analysis

ssGSEA

Single-sample gene set enrichment analysis

BLCA

Bladder urothelial carcinoma

RMSD

Root mean square deviation

RMSF

Root mean square fluctuation

MFH

Medicine-food homology

TCMSP

Traditional Chinese Medicine Systems Pharmacology Database

SMILES

Simplified Molecular Input Line Entry System

IHC

Immunohistochemistry

IF

Immunofluorescence

CCK-8

Cell Counting Kit-8

CAFs

Cancer-associated fibroblasts

FGF

Fibroblast growth factor

MIF

Macrophage migration inhibitory factor

LAMs

Lipid-associated macrophages

TAMs

Tumor-associated macrophages

Authors’ contributions

SRC, RSS and AQL designed the study. SRC conducted the bioinformatic analysis and analyzed the data. RSS, ZLY and HHC performed the experiments and analyzed the data. SRC and RSS drafted the manuscript. ZLY, HHC, AQL and YYL made a significant contribution to the acquisition of the data. AQL and YYL reviewed and revised the manuscript. All authors contributed to the article and approved the submitted version.

Funding

This study was supported by the Application Medical Research Project of Hefei Municipal Health Commission (Grant No.Hwk2024qlyb001).

Data availability

Publicly available datasets were analyzed in this study. These data can be found in https://gdc.xenahubs.net/, https://www.cbioportal.org/, http://research-pub.gene.com/IMvigor210CoreBiologies/, http://swisstargetprediction.ch/.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval

This study was conducted in accordance with the Declaration of Helsinki. The clinicopathological data and BC tissues used were obtained from patients who underwent surgical treatment at the First People’s Hospital of Hefei. This study was approved by the Ethics Committee of the First People’s Hospital of Hefei [registration number: 2024—049—01]. Informed consent was obtained from all individual participants included in the study.

Consent for publication

Not applicable.