Integrated bulk and single-cell transcriptomic analysis reveals a tryptophan metabolism-driven prognostic signature and therapeutic landscape in triple- negative breast cancer

Integrated bulk and single-cell transcriptomic analysis reveals a tryptophan metabolism-driven prognostic signature and therapeutic landscape in triple- negative breast cancer | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Integrated bulk and single-cell transcriptomic analysis reveals a tryptophan metabolism-driven prognostic signature and therapeutic landscape in triple- negative breast cancer Youjun Wu, Xiaorong Pang, Feng Cen, Liang Xie, Xiang Feng, Xianglan Mo This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9287896/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background: Triple-negative breast cancer (TNBC) is an aggressive malignancy with limited therapeutic options and poor prognosis. The kynurenine pathway of tryptophan metabolism contributes to an immunosuppressive tumor microenvironment (TME); however, its prognostic significance and molecular mechanisms in TNBC require further multi-omics characterization. Methods: Bulk and single-cell RNA-seq datasets were obtained from public repositories. Differentially expressed tryptophan metabolism-related genes were identified and screened via univariate Cox regression. Nine machine learning algorithms were trained using 5-fold cross-validation, with SHAP analysis applied for model interpretability. A prognostic risk model was developed, externally validated, and further analyzed for immune infiltration, pathway activity, and AI-driven drug screening. ScRNA-seq data were used to identify key cell populations and differentiation trajectories. Results: Among 48 candidates, Random Survival Forest demonstrated optimal performance and was selected to construct an eight-gene prognostic signature (EIF4EBP1, NRTN, COL9A3, TRIM63, FABP7, ALAD, H4C13, PLAU). The model effectively stratified patients into high- and low-risk groups with significantly distinct survival outcomes across multiple cohorts. High-risk patients exhibited increased infiltration of central memory CD8+ T cells, immature dendritic cells, and neutrophils, along with upregulated ABC transporter and endocytosis pathways. AI-driven screening identified omega-3-carboxylic acids and sucralfate as potential therapeutics. ScRNA-seq revealed that prognostic markers were predominantly expressed in T cells, B cells, macrophages, and stromal cells, with dynamic changes along differentiation trajectories. Conclusion: This study establishes a novel tryptophan metabolism-derived prognostic signature for TNBC, providing insights into TME remodeling and identifying potential therapeutic strategies for high-risk patients. Triple-negative breast cancer Tryptophan metabolism Machine learning Prognostic model Single-cell RNA sequencing Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 1 Introduction Breast cancer, the second most common cancer worldwide after lung cancer in terms of incidence, is the leading cause of cancer mortality(Bray et al. 2024 ). Triple-negative breast cancer (TNBC) is defined by the absence of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2)(Dowling et al. 2021). This unique molecular profile renders it refractory to endocrine therapy and anti-HER2 targeted agents(Wang et al. 2020 ), which consequently establishes TNBC as the breast cancer subtype with the most aggressive clinical behavior and poorest prognosis. TNBC comprises 10–20% of all breast cancers and is distinguished by a set of clinical features, including increased aggressiveness, elevated recurrence risk, and a proclivity for visceral dissemination(Zhou et al. 2021 , Zhang et al. 2022b). The therapeutic landscape for TNBC remains dominated by cytostatic chemotherapy, yet the benefits of this treatment have reached a plateau in improving patient outcomes(Chang et al. 2019). Although emerging modalities, notably immune checkpoint inhibitors, offer renewed hope for a subset of patients, their clinical application is hampered by limited response rates and a critical lack of predictive biomarkers to reliably identify beneficiaries(Ji et al. 2022). Consequently, deciphering the fundamental molecular drivers of TNBC to develop reliable prognostic tools and identify new therapeutic vulnerabilities is an urgent and critical research imperative. Metabolic reprogramming, a core cancer hallmark proposed by Hanahan (Hanahan 2022) supports tumor growth and fosters an immunosuppressive microenvironment. The tryptophan-kynurenine pathway is a critical mediator of this metabolic reprogramming process(Yan et al. 2024, Xue et al. 2023a). In TNBC, up to 95% of tryptophan is metabolized via this route(Lin et al. 2025). The enzymes indoleamine 2,3-dioxygenase 1 (IDO1) and tryptophan 2,3-dioxygenase 2 (TDO2) produce kynurenines, which impair T-cell function through two distinct mechanisms: by causing local tryptophan depletion and, via activation of the aryl hydrocarbon receptor (AhR), promoting regulatory T cell (Treg) differentiation while inhibiting effector T cells, thereby enabling tumor immune evasion(Kuo et al. 2024, Rogers et al. 2019, Greene et al. 2019). The clinical failure of IDO1 inhibitors reveals metabolic network complexity and redundancy, which demands a systems-level understanding of the regulation of tryptophan metabolism and inter-pathway crosstalk(Lercher et al. 2020, Jahchan et al. 2019). Traditional prognostic models based on bulk RNA sequencing, while informative of overall gene expression, are limited by their inability to dissect intratumoral cellular complexity and interactions(Tan et al. 2023 ). Single-cell RNA sequencing (scRNA-seq) overcomes this by providing high-resolution maps of the tumor microenvironment and tracing key events to specific cell types(Ma et al. 2023). However, scRNA-seq data alone is often insufficient for building robust clinical prognostic models. Thus, a combined approach—leveraging the prognostic power of bulk RNA-seq data and the cellular resolution of single-cell RNA-seq (scRNA-seq) data—represents a more powerful research strategy. The "black-box" nature of machine learning prognostic models impedes their clinical adoption. Explainable AI (XAI) approaches like SHapley Additive exPlanations (SHAP) mitigate this by quantifying feature-specific contributions, thus preserving predictive power while enabling biological interpretation(Uthayopas et al. 2021, Yu et al. 2022). In a parallel challenge within translational medicine, the inefficient conversion of basic research findings into clinical therapies remains a major obstacle. AI-based drug discovery models, including Graph Neural Networks (GNNs) like GraphBAN, are addressing this by systematically screening for target-interacting compounds, accelerating both drug repurposing and novel drug development(Hadipour et al. 2025). This study employs an integrated multi-omics and computational strategy to define the prognostic value of tryptophan metabolism in TNBC. Our methodology involved constructing a prognostic model based on bulk RNA-seq transcriptomic data, followed by SHAP-based analysis to elucidate the contribution of key genes to the model's predictions. Furthermore, we will leverage artificial intelligence-based drug prediction platforms to identify potential lead compounds targeting this pathway. Finally, scRNA-seq data was utilized to dynamically resolve the cell-type-specific mechanisms underlying these associations between tryptophan metabolism and prognosis at the microscopic level. The overarching goal of this research is to provide novel therapeutic targets and a conceptual framework to advance precision medicine for TNBC. 2 Materials and methods 2.1 Data source On July 21, 2024, we retrieved bulk RNA-sequencing data along with clinical characteristics for 100 TNBC samples and 113 normal breast tissue specimens from The Cancer Genome Atlas (TCGA) repository ( https://portal.gdc.cancer.gov/ ), designated as TCGA-TNBC cohort. For external validation purposes, we obtained the GSE135565 dataset (GPL570 platform) from the Gene Expression Omnibus (GEO) database ( https://www.ncbi.nlm.nih.gov/geo/ ), which contained 83 TNBC cases with corresponding survival outcomes. Additionally, scRNA-seq data (GSE161529, GPL18573 platform) encompassing 4 TNBC and 13 normal breast tissue specimens were acquired from GEO. Genes associated with tryptophan metabolism (TMRGs) were compiled from dual database sources on July 21. Initially, we searched GeneCards ( https://www.genecards.org ) using "tryptophan metabolism" as the query term, yielding 4,951 protein-coding genes. From this pool, we retained 3,113 genes demonstrating a relevance score ≥ 10. Subsequently, we extracted 51 genes from the Molecular Signatures Database (MSigDB) ( https://www.gsea-msigdb.org ) across three gene collections: "KEGG_TRYPTOPHAN_METABOLISM", "REACTOME_TRYPTOPHAN_CATABOLISM", and "WP_TRYPTOPHAN_METABOLISM". Following consolidation and duplicate removal, we established a comprehensive TMRG collection comprising 3,116 genes. 2.2 Identification of candidate genes Using the DESeq2 package (v 1.40.2)(Love et al. 2014), we performed differential expression profiling between TNBC and control specimens within the TCGA-TNBC cohort. Genes exhibiting adjusted p-values (adj.p.val) 1 were classified as differentially expressed genes (DEGs). Visualization of DEGs was accomplished through volcano plots generated by ggplot2 package (v 3.4.1) (Gustavsson et al. 2022) and heatmaps created using ComplexHeatmap package (v 2.14.0)(Gu and Hübschmann 2022 ). The top 10 upregulated and downregulated genes, ordered by descending |log 2 FC|, were annotated in volcano plots, with their expression profiles further illustrated in heatmaps. We identified differentially expressed TMRGs (DE-TMRGs) through intersection of DEGs with the TMRG collection using VennDiagram package (v 1.7.1)(Chen and Boutros 2011 ). To pinpoint DE-TMRGs significantly correlated with overall survival (OS), we employed univariate Cox proportional hazards regression via the survival package (v 3.7.0) (Lei et al. 2023)(hazard ratio (HR) ≠ 1, p 0.05 being selected as candidate genes. 2.3 Functional enrichment and protein-protein interaction (PPI) analysis of candidate genes To decipher the biological roles of selected candidate genes, we conducted Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment evaluations using clusterProfiler package (v 4.2.2)(Wu et al. 2021a ). (adj.p.val < 0.05) GO analysis encompassed three ontological dimensions: biological processes (BPs), cellular components (CCs), and molecular functions (MFs). The top 5 pathways ranked by adj.p.val from each GO domain and KEGG pathways were visualized via ggplot2 package (v 3.4.1). To explore protein-level interactions among candidate genes, we constructed a protein-protein interaction (PPI) network utilizing the Search Tool for the Retrieval of Interacting Genes (STRING) database ( https://cn.string-db.org/ ) with a minimum interaction score threshold of 0.15. 2.4 Machine learning model construction and SHapley additive exPlanations (SHAP) analysis We developed nine distinct machine learning algorithms using candidate gene data and OS information from TCGA-TNBC: Cox proportional hazards (coxph), survival support vector machine (survivalsvm), stepwise Cox regression (Stepwise), elastic net, Ridge regression, Lasso regression, random survival forest (RandomForestSRC), partial least squares Cox regression (plsRcox), and extreme gradient boosting (xgboost). Algorithm performance was assessed through 5-fold cross-validation procedures. Performance indicators included concordance index (C-index), time-dependent area under the curve (AUC) at 1, 2, and 3 years, and Brier scores. The algorithm demonstrating the highest C-index alongside AUC values exceeding 0.6 was designated as optimal. We implemented SHAP analysis using shapviz package (v 0.9.7) (Sugiawan et al. 2023 ) to interpret the optimal algorithm's predictions, generating multiple visualization formats including bar plots, beeswarm plots, dependence plots, waterfall plots, and force plots. 2.5 Prognostic model construction and validation We constructed a random survival forest (RSF) algorithm using randomForestSRC package (v 3.3.1) (Jin et al. 2024)based on candidate genes from the TCGA cohort. Genes demonstrating importance scores > 0.025 were designated as final prognostic markers. The algorithm's predictive output served as the risk score. Patient stratification into high-risk group (HRG) and low-risk group (LRG) was achieved using the optimal cut-point determined by the surv_cutpoint function from survminer package (v 0.4.9) (Liu et al. 2021)with minprop = 0.4. Kaplan-Meier (K-M) survival curves with log-rank tests evaluated survival disparities between groups (p < 0.05). Furthermore, the expression heatmap of prognostic markers, risk score distribution, and patient survival status across HRG and LRG were comprehensively visualized using ggplot2 package (v 3.4.1). Predictive accuracy of the risk score was determined by calculating the area under the ROC curve (AUC) for 1, 2, and 3-year OS using survivalROC package (v 1.0.3.1)(Heagerty et al. 2000). This entire methodology was validated in the independent GSE135565 cohort. 2.6 Independent prognostic analysis Within TNBC specimens with available survival data from TCGA-TNBC, we performed univariate and multivariate Cox regression analyses using survival package (v 3.7.0) to evaluate the independence of risk scores and clinical parameters (age, tumor stage, etc.). Variables exhibiting HR ≠ 1, p 0.05) were deemed independent prognostic factors. 2.7 Immune infiltration, gene set variation analysis (GSVA) , and gene set enrichment analysis (GSEA) We utilized GSVA package (v 1.50.0) (Hänzelmann et al. 2013 ) with the single-sample gene set enrichment analysis (ssGSEA) algorithm to quantify relative abundances of 28 immune cell subsets (Charoentong et al. 2017) within each TNBC specimen from TCGA-TNBC dataset. Disparities in immune cell infiltration between risk categories were evaluated using Wilcoxon rank-sum test (p < 0.05). Correlations between prognostic gene expression/risk scores and significantly altered immune cells were computed using ggpubr package (v 0.6.0) (Cheng et al. 2021 )(|correlation coefficient (cor)| > 0.3, p < 0.05). To investigate underlying biological pathways exhibiting differential activation between HRG and LRG, we executed GSVA using GSVA package (v 1.42.0)(Hänzelmann et al. 2013 ). The reference gene collection employed was "c2.cp.kegg.v2023.1.Hs.symbols.gmt", a curated compilation of KEGG database pathways, obtained from MSigDB ( http://www.gsea-msigdb.org/gsea/msigdb/index.jsp ). The ssGSEA algorithm computed enrichment scores for each pathway across every TNBC specimen in the TCGA-TNBC cohort. Subsequently, differential pathway activity between HRG and LRG was determined using limma package (v 3.54.0) (Ritchie et al. 2015)(|t| > 2, p < 0.05). To further explore the biological pathway differences between HRG and LRG, GSEA was performed using the "c2.cp.kegg.v2023.1.Hs.symbols.gmt" gene set from the MSigDB. Differential expression analysis between risk groups was conducted using the DESeq2 package (v 1.40.2), and genes were ranked by log 2 FC. GSEA was implemented using the GSEA function in the clusterProfiler package (v 4.2.2), with significance thresholds set at p 1. 2.8 Tumor mutation burden (TMB) analysis Somatic mutation profiles for TNBC specimens from the TCGA cohort were retrieved and processed via TCGAmutations package (v 0.3.0)(Tan et al. 2025 ). Mutation data were analyzed and visualized using maftools package (v 2.16.0)(Zhang et al. 2023 ). TMB was computed as the total mutation count per megabase. TMB differences between risk categories were assessed using Wilcoxon test. TNBC patients across the entire cohort were stratified into high-TMB and low-TMB categories based on median TMB values. K-M survival curves were subsequently generated for these categories, with statistical significance of OS differences evaluated using log-rank test (p < 0.05). 2.9 AI-based drug prediction To translate prognostic genes into therapeutic opportunities, we aimed to identify small-molecule compounds capable of binding proteins encoded by prognostic genes. The objective centered on computational screening for high-affinity interactions, establishing rationale for drug repurposing or novel therapeutic development targeting the identified prognostic pathway in TNBC. We employed the Graph-Based Attention Network (GraphBAN) algorithm to predict compound-protein interactions (CPIs). SMILES notations of compounds were sourced from ChEMBL database ( https://www.ebi.ac.uk/chembl/ ), while protein sequences were retrieved from UniProt ( https://www.uniprot.org/ ). Molecular graph features of compounds were extracted using ChemBERTa and Graph Convolutional Networks (GCN). Protein embedding features were generated through 1D-CNN and Evolutionary Scale Modeling (ESM). These features were fed into the GraphBAN algorithm, which utilizes bidirectional attention mechanisms to predict interaction probabilities. Compound-protein pairs demonstrating interaction probabilities > 0.8 were considered high-affinity candidates. Network visualization was accomplished using Cytoscape (v 3.10.2)(Liu et al. 2020). 2.10 Drug sensitivity analysis To assess the association between the prognostic risk model and chemotherapeutic response, drug sensitivity data were obtained from the Cancer Therapeutics Response Portal (CTRP) and the Genomics of Drug Sensitivity in Cancer (GDSC) database ( https://www.cancerrxgene.org/ ). The half-maximal inhibitory concentration (IC 50 ) of 138 common chemotherapeutic and molecular targeted agents was estimated using the pRRophetic package (v 0.5), which employs a ridge regression model to predict drug response based on pre-treatment tumor gene expression profiles. IC 50 values were compared between HRG and LRG using the Wilcoxon test (p < 0.05). 2.11 scRNA-seq analysis We employed Seurat package (v 5.1.0) (Satija et al. 2015) for scRNA-seq data processing. Quality control involved filtering low-quality cells (nFeature_RNA 8000, mitochondrial gene percentage > 10%). Data normalization was performed using NormalizeData function. The top 3000 highly variable genes (HVGs) were identified via FindVariableFeatures function. Principal component analysis (PCA) was executed, with optimal principal component (PC) numbers selected for downstream analysis based on elbow plots and JackStraw tests. Cell clustering was performed using FindNeighbors and FindClusters functions (resolution = 2) and visualized through Uniform Manifold Approximation and Projection (UMAP). Cell type annotation was accomplished using SingleR package (v 2.2.0) (Aran et al. 2019)and manually refined with CellMarker database ( https://ngdc.cncb.ac.cn/cellmarker/ ). Differential abundance of cell types between TNBC and control specimens was evaluated using Wilcoxon test (p < 0.05). Key cell types were defined as those exhibiting differential abundance and significant differential expression of prognostic genes (p < 0.05, Wilcoxon test). 2.12 Pseudotime, cell-cell communication, transcription factor (TF), and metabolic activity analysis Pseudotime trajectory inference was conducted using Monocle2 package (v 2.28.0)(Cao et al. 2019). Genes exhibiting dynamic expression patterns along pseudotime were identified and visualized. Transcription factor (TF) activity was inferred using decoupleR package (v 2.6.0)(Veghini et al. 2024). Metabolic pathway activity was quantified using scMetabolism package (v 0.2.1)(Zhang et al. 2022a). Dot plots visualized metabolic activity patterns across key cell types. To investigate intercellular communication networks within the tumor microenvironment, cell-cell interaction analysis was performed using the CellChat package (v 1.6.1) on the scRNA-seq dataset. This algorithm quantitatively infers and analyzes intercellular communication networks by integrating ligand-receptor interaction databases with gene expression data. The number and strength of intercellular interactions were calculated for each cell type, and communication probabilities were estimated using a statistical model. Signaling networks were visualized for both TNBC and control samples. 2.13 Statistical analysis All statistical analyses were conducted using R software (v 4.2.2). Unless stated otherwise, Wilcoxon test was employed to assess between-group differences with significance threshold set at p < 0.05. 3 Results 3.1 Identification and functional enrichment of tryptophan metabolism-related candidate genes A total of 6,136 DEGs were identified between TNBC and control tissues, including 3,656 upregulated and 2,480 downregulated genes (adj.p.val 1) (Fig. 1 A-B, Supplementary Table S1 ). Among these, 1,058 overlapped with TMRGs and were designated DE-TMRGs (Fig. 1 C). Univariate Cox regression analysis identified 54 DE-TMRGs significantly associated with OS (HR ≠ 1, p 0.05) (Fig. 1 E). After removing one gene (CYP11B1) with an exceptionally large HR (> 10,000), 48 genes were finalized as candidate genes for subsequent analysis. GO and KEGG enrichment analyses were performed to elucidate the potential biological functions and signaling pathways associated with the 48 candidate genes. The results indicated that these genes were significantly enriched in BPs critical for tumor progression, including ''response to estradiol'', ''stem cell development'', and ''response to oxygen levels'' (adj.p.val < 0.05) (Fig. 1 F). This enrichment pattern suggests that the candidate genes may orchestrate a pro-tumorigenic program by modulating hormone sensitivity, maintaining cancer stemness, and adapting to hypoxic stress within the tumor microenvironment (TME). In terms of MFs and CCs, the genes were predominantly associated with ''extracellular matrix (ECM) structural constituent'' and ''collagen-containing ECM'', highlighting their potential role in remodeling the ECM to facilitate tumor invasion, metastasis, and cell-matrix communication. KEGG pathway analysis further corroborated these findings, revealing significant enrichment in key oncogenic pathways, most notably the ''PI3K-Akt signaling pathway'' and ''ECM-receptor interaction'' (adj.p.val < 0.05) (Fig. 1 G). The enrichment of the PI3K-Akt pathway, a well-established driver of cell survival, proliferation, and metabolism in TNBC, implies that our candidate gene signature may contribute to tumor aggressiveness and therapy resistance through this axis. Concurrently, the enrichment in ECM-receptor interaction underscores a mechanistic link to enhanced cell adhesion, migration, and activation of integrin-mediated survival signals. To validate these functional associations at the protein level, a PPI network was constructed. The network, comprising 47 nodes and 254 edges, exhibited robust biological connectivity among the candidate genes (Fig. 1 H). This highly interconnected network suggests that these proteins do not function in isolation but rather as coordinated modules or complexes. 3.2 Construction and interpretation of the optimal model Nine machine learning models were trained and evaluated via 5-fold cross-validation. The RSF model demonstrated the highest C-index and maintained AUCs > 0.6 across 1, 2, and 3 years, establishing it as the optimal model for prognosis prediction (Table 1 ). SHAP analysis was applied to interpret the RSF model. The bar plot ranked the importance of the candidate genes, with NRTN and EIF4EBP1 showing high impacts (Fig. 2 A). A beeswarm plot summarized the distribution of SHAP values for each gene across all individual patients in the TCGA cohort (Fig. 2 B). The plot revealed that high expression of EIF4EBP1 (orange points clustered on the positive SHAP value side) was consistently associated with an increased risk prediction, whereas high expression of NRTN (orange points clustered on the negative SHAP value side) was associated with a decreased risk prediction. Notably, NRTN showed a wide dispersion of SHAP values, indicating its effect size varied considerably between patients. Dependence plots for the 4 most important genes demonstrate that increasing expression of EIF4EBP1 increased the predicted risk, while increasing expression of NRTN, FABP7, and COL9A3 decreased it (Fig. 2 C). This reinforced their roles as potential risk and protective factors, respectively. Table 1 Selecting the optimal machine learning model Model C_index AUC_1yr AUC_2yr AUC_3yr Brier_1yr Brier_2yr Brier_3yr coxph ★ 0.885 ± 0.117 0.022 ± 0.038 0.043 ± 0.051 0.197 ± 0.382 0.525 ± 0.093 0.535 ± 0.022 0.545 ± 0.091 rfsrc ★ 0.800 ± 0.169 0.705 ± 0.418 0.897 ± 0.179 0.827 ± 0.208 0.330 ± 0.043 0.488 ± 0.115 0.548 ± 0.085 ridge 0.784 ± 0.184 0.787 ± 0.084 0.939 ± 0.105 0.864 ± 0.180 0.392 ± 0.048 0.593 ± 0.143 0.678 ± 0.112 xgboost 0.768 ± 0.228 0.773 ± 0.321 0.835 ± 0.164 0.794 ± 0.179 0.251 ± 0.025 0.308 ± 0.086 0.334 ± 0.083 stepwise 0.682 ± 0.251 NaN ± NA 0.833 ± NA 0.900 ± 0.141 0.770 ± 0.315 0.840 ± 0.219 0.870 ± 0.179 plsRcox 0.660 ± 0.255 0.510 ± 0.178 0.732 ± 0.330 0.683 ± 0.399 0.444 ± 0.068 0.459 ± 0.166 0.431 ± 0.156 lasso 0.575 ± 0.107 0.795 ± 0.289 0.807 ± 0.269 0.608 ± 0.193 0.306 ± 0.077 0.350 ± 0.159 0.349 ± 0.168 elastic 0.545 ± 0.077 0.545 ± 0.064 0.559 ± 0.261 0.574 ± 0.220 0.279 ± 0.064 0.349 ± 0.173 0.361 ± 0.186 survivalsvm 0.500 ± 0.000 0.168 ± 0.020 0.118 ± 0.102 0.118 ± 0.102 0.410 ± 0.042 0.620 ± 0.135 0.710 ± 0.096 Waterfall and force plots for individual samples demonstrated how each gene's expression pushed the model's prediction towards a high or low risk score (Fig. 2 D). The waterfall plot showed the model predicts a high-risk score of f(x) = 9.13 for this sample, significantly above the baseline expected value E [f(x)] = 2.63. The explanation revealed that 43 genetic features collectively contribute to this elevated risk. The most substantial positive contributions come from a subset of features, with EIF4EBP1 contributing + 8.83 SHAP value units. These large positive values indicated that the expression pattern of the specific genes in this sample was the primary driver of its high-risk classification. 3.3 Development and validation of a robust prognostic model based on prognostic genes The RSF model with the 48 candidate genes was used to calculate a risk score. Eight genes with an importance score > 0.025 (EIF4EBP1, NRTN, COL9A3, TRIM63, FABP7, ALAD, H4C13, PLAU) were identified as the prognostic genes. Patients in both the TCGA training and GSE135565 validation cohorts were stratified into HRG and LRG using the optimal cut-points (TCGA-TNBC: 3.109546, GSE135565: 5.751818). K-M analysis confirmed that patients in the HRG had significantly worse OS than those in the LRG in both cohorts (log-rank p < 0.05) (Fig. 3 A). The distribution of risk scores, survival status, and expression heatmap of the eight prognostic genes further validated the model's stratification power (Fig. 3 B). Time-dependent ROC analysis showed that the risk score had predictive value for 1-, 2-, and 3-year OS, with AUCs consistently above 0.8 in both cohorts (Fig. 3 C). 3.4 Risk score as an independent prognostic factor Univariate Cox regression identified the risk score, N stage, and M stage as potential prognostic factors (Cox p 0.05) (Fig. 4 A-B). Subsequent multivariate Cox regression confirmed that the risk score and M stage were independent prognostic factors for OS in TNBC (Cox p 0.05) (Fig. 4 C-D). These results underscore the clinical relevance of the TMRG-based signature beyond conventional clinicopathological variables. 3.5 Immune infiltration and pathway activity in risk groups The ssGSEA revealed significant differences in immune cell infiltration between risk groups (Fig. 5 A-B). Central memory CD8 + T cells, immature dendritic cells, and neutrophils were significantly enriched in the HRG (p 0.3, p < 0.01), while NRTN was significantly negatively correlated with central memory CD8 T cells and neutrophils (cor < -0.4, p < 0.01) (Fig. 5 C). The co-enrichment of these immune cells, coupled with their divergent correlations with PLAU and NRTN, suggests a complex immunomodulatory role for these prognostic genes, potentially contributing to an immunosuppressive microenvironment in high-risk TNBC. GSVA highlighted significant pathway activity differences between risk groups. Seven KEGG pathways were significantly enriched (|t| > 2, p < 0.05) (Fig. 5 D). Among these, five pathways were significantly activated in the HRG, including ''ABC transporters'', ''N glycan biosynthesis'', ''type II diabetes mellitus'', ''riboflavin metabolism'', and ''endocytosis''. Conversely, pathways including ''cysteine and methionine metabolism'' and ''glycosaminoglycan biosynthesis keratan sulfate'' were suppressed. The specific activation of ABC transporters and endocytosis in high-risk TNBC reveals a coordinated mechanism for drug efflux and membrane trafficking that likely contributes to chemoresistance, while the altered glycosylation and metabolic patterns suggest extensive reprogramming of cell surface properties and redox homeostasis that collectively promote tumor survival and progression. To elucidate the biological pathways distinguishing high- from low-risk TNBC patients, GSEA was performed using KEGG gene sets. A total of 19 pathways were significantly enriched (p 1). The top five significantly enriched pathways in the HRG included ''steroid hormone biosynthesis'', ''pentose and glucuronate interconversions'', ''drug metabolism-other enzymes'', ''retinol metabolism'', and ''ascorbate and aldarate metabolism'' (Fig. 5 E). These findings indicate that high-risk TNBC is characterized by pronounced metabolic reprogramming involving hormone metabolism, xenobiotic processing, and oxidative stress-related pathways, which may contribute to its aggressive phenotype and therapeutic resistance. 3.6 Somatic mutational landscape and prognostic value of TMB The waterfall diagram illustrated the somatic mutational profile across 91 TNBC samples, with 88 samples (96.7%) exhibiting at least one genetic alteration (Fig. 6 A). TP53 demonstrated the highest mutation frequency, present in approximately 84% of samples, followed by TTN (19%) and MUC16 (14%). The mutation spectrum was dominated by missense mutations across all major genes. No significant difference in TMB was observed between risk groups (p = 0.858) (Fig. 6 B). However, high TMB was associated with improved OS (p = 0.032) (Fig. 6 C). Patients with low risk and high TMB had the most favorable prognosis (p < 0.0001) (Fig. 6 D), highlighting the complementary value of genomic and transcriptomic biomarkers. 3.7 AI-predicted therapeutic compounds GraphBAN analysis predicted high-affinity interactions between four prognostic genes (H4C13, COL9A3, ALAD, TRIM63) and two compounds: omega-3-carboxylic acids (ω-3 carboxylic acid) and sucralfate (interaction probability > 0.8) (Table 2 , Fig. 7 A). Among the prognostic genes, COL9A3 was found to have the highest number of connecting edges, suggesting its higher potential as a novel drug target. Conversely, among the compounds, ω-3 carboxylic acid exhibited the most connections, indicating their strong potential as a novel therapeutic agent. No corresponding compounds were predicted for the remaining four prognostic genes. These findings suggest repurposing opportunities for existing drugs in TNBC treatment. Table 2 Prediction results from GraphBAN SMILES Protein pred Drug Name Gene CC/...O)O MSG...GGF 0.800178885 OMEGA-3-CARBOXYLIC ACIDS H4C13 CC/...O)O MAG...RSS 0.815427303 OMEGA-3-CARBOXYLIC ACIDS COL9A3 CC/...O)O MQP...KEE 0.889059484 OMEGA-3-CARBOXYLIC ACIDS ALAD CC/...O)O MDY...GHQ 0.823631704 OMEGA-3-CARBOXYLIC ACIDS TRIM63 O.O...O)O MAG...RSS 0.806707144 SUCRALFATE COL9A3 3.8 Different drug sensitivity between HRG and LRG To evaluate the potential clinical utility of the prognostic model in guiding treatment selection, the IC 50 values of 138 chemotherapeutic and targeted agents were estimated in the TCGA-TNBC cohort. Wilcoxon test identified ten drugs with significantly different IC 50 values between risk groups (p < 0.05). Among these, the majority exhibited lower IC 50 values in the LRG, including AZ628, indicating greater drug sensitivity (Fig. 7 B). These results suggest that patients in the LRG may be more responsive to conventional chemotherapeutic agents, highlighting the potential utility of the prognostic model in refining treatment strategies. 3.9 Identification of key cells After quality control ( Supplementary Fig. 1A ), the top 3,000 HVGs and 20 PCs were selected for cell clustering ( Supplementary Fig. 1B-D ). Unsupervised clustering with resolution parameter 2 identified 53 distinct cell clusters (Fig. 8 A), which were annotated into 8 major cell types: B cells, common myeloid progenitors (CMPs), dendritic cells (DCs), endothelial cells, epithelial cells, macrophages, stromal cells, and T cells (Fig. 8 B). Differential abundance analysis revealed six cell types with significantly different abundances between TNBC and control tissues (p < 0.05) (Fig. 8 C). Stromal cells were significantly decreased in TNBC, while five immune cell types (T cells, B cells, macrophages, DCs, CMPs) were enriched, indicating immune activation in the TNBC microenvironment. UMAP visualization showed distinct expression patterns for prognostic genes (Fig. 8 D). Subsequently, differential expression analysis showed T cells, B cells, macrophages, and stromal cells exhibited significant differential expression of at least one prognostic gene and were therefore defined as the four key cell types for further analysis (Fig. 8 E). DCs showed no significant differences, and CMPs were excluded as they were only present in TNBC samples. 3.10 Subcluster analysis and pseudotemporal trajectory of key cell types Subsequent secondary dimensionality reduction and clustering were performed on the four key cell types (macrophages, T cells, B cells, and stromal cells). Macrophages were subdivided into 14 distinct subclusters, T cells into 12 subclusters, B cells into 14 subclusters, and stromal cells into 9 subclusters (Fig. 9 A), revealing substantial heterogeneity within each major cellular compartment. Pseudotemporal trajectory analysis was further conducted to reconstruct the potential differentiation pathways or cellular state transitions within each of these key cell types. Each cell type exhibited a unique branched trajectory, visualized as a continuum from a presumed initial state (dark blue) towards a terminal state (light orange) along the inferred pseudotime (Fig. 9 B). Analysis of genes demonstrating significant dynamic expression changes during these transitions highlighted key mediators specific to each lineage. Notably, CCL20 in macrophages, CCL4 in T cells, FCER1G in B cells, and ACTG2 in stromal cells showed pronounced expression variations along their respective trajectories, as illustrated in expression heatmaps and trend plots (Fig. 9 C-D). The dynamic expression patterns of the prognostic genes were specifically examined along these pseudotemporal axes (Fig. 9 E). Distinct temporal expression programs were observed. In macrophages, FABP7 expression was biased towards the beginning of the pseudotime, whereas PLAU expression was enriched towards the end. For T cells, FABP7 exhibited higher expression during the early-to-mid phases of the differentiation trajectory. During B-cell differentiation, both FABP7 and PLAU showed elevated expression in the early phase. Stromal cells displayed a more complex pattern, with dynamic expression observed for EIF4EBP1, COL9A3, FABP7, and PLAU. Specifically, EIF4EBP1 and PLAU maintained relatively high expression throughout the early, middle, and late stages, COL9A3 was highly expressed in the early phase, and FABP7 expression increased during the later phase. These findings delineate cell-type-specific and differentiation-stage-dependent expression dynamics of the prognostic genes within the TNBC tumor microenvironment, suggesting their potential roles in distinct biological processes across different cellular contexts. 3.11 TF activity and metabolic pathway activity The TF activity analysis provided crucial insights into the regulatory programs operating in key cells of TNBC (Fig. 10 A). The consistent activation of MYC and E2F4, coupled with suppression of ATF3 and HBP1, suggested coordinated pro-tumorigenic regulatory networks across different cell types. Metabolic scoring indicated elevated activity in pathways such as glycolysis/gluconeogenesis and fatty acid elongation in macrophages compared to other cell types (Fig. 10 B). This aligns with the known metabolic plasticity of tumor-associated macrophages in TNBC. 3.12 Intercellular communication networks In TNBC samples, B cells, T cells, and macrophages showed enhanced interaction intensity compared with controls, whereas stromal cells displayed stronger interactions in control tissues (Fig. 11 A-B). Analysis of ligand-receptor pair contributions revealed that macrophage-derived SPP1 signaling via the CD44 receptor complex was specifically activated in TNBC (Fig. 11 A-B). Collectively, these findings reveal a rewired intercellular communication landscape in TNBC, with active participation of immune and stromal cells in tumor-promoting signaling networks. 4 Discussion Triple-negative breast cancer (TNBC) is an aggressive malignancy with limited therapies, poor prognosis, and a lack of reliable biomarkers. Tryptophan metabolism—particularly the kynurenine pathway—can promote an immunosuppressive tumor microenvironment (TME), though its prognostic role in TNBC remains unclear. Using multi-omics data, we identified 48 tryptophan metabolism-related genes by differential expression and univariate Cox analysis. An optimized random survival forest (RSF) model was built and validated as an 8-gene prognostic signature, including EIF4EBP1 and NRTN, which SHAP analysis highlighted as key risk stratification factors. High-risk TNBC showed an "infiltrated but suppressed" immune profile, with enriched central memory CD8⁺ T cells, immature dendritic cells, and neutrophils, plus activated oncogenic pathways like ABC transporters and endocytosis. The GraphBAN AI model predicted ω-3 carboxylic acid and sucralfate as potential therapeutics. scRNA-seq revealed cell-specific expression and differentiation trajectories of prognostic genes in T cells, B cells, macrophages, and stromal cells. This study offers a tryptophan metabolism-derived prognostic tool for TNBC, clarifies TME remodeling mechanisms, and suggests translational drug candidates to advance precision oncology. A key finding is the development and validation of a robust prognostic model based on eight genes (EIF4EBP1, NRTN, COL9A3, TRIM63, FABP7, ALAD, H4C13, PLAU). The model showed strong predictive performance in both the TCGA training and GEO validation sets, and its risk score independently predicted overall survival and surpassed certain conventional TNM staging criteria, thereby highlighting its added clinical value, and these results imply that tryptophan metabolic reprogramming plays a more central role in TNBC progression than previously appreciated, reflecting tumor intrinsic aggressiveness(Yan et al. 2024, Xue et al. 2023a). Using machine learning, Random Survival Forest (RSF) was selected as the optimal model for its predictive accuracy and ability to capture complex gene–gene interactions, consistent with the polygenic basis of tumorigenesis. SHAP analysis enhanced interpretability, identifying EIF4EBP1 and NRTN as the top contributors to high- and low-risk stratification, respectively, offering clear mechanistic targets; the high-risk gene EIF4EBP1, a downstream effector of mTORC1, regulates protein translation and cancer metabolism(Wu et al. 2024 , Fang et al. 2022). Under nutrient-rich conditions, mTORC1 phosphorylates and inhibits EIF4EBP1, promoting oncogenic protein synthesis. Consistent with prior reports(Wu et al. 2024 ), high EIF4EBP1 expression correlated with poor prognosis, possibly indicating sustained mTOR activation or non-canonical pro-tumor functions(Guo et al. 2020). In contrast, In contrast, NRTN, a GDNF family member mainly linked to neuronal survival(Correia et al. 2021), has an unclear role in TNBC; we propose that it may interact via GFRα2 with integrins on tumor or immune cells, potentially modulating ECM remodeling and pro-invasive signaling. Although it has an unclear role in TNBC, and we propose that NRTN it may interact via GFRα2 with integrins on tumor or immune cells to, potentially modulating ECM remodeling and pro-invasive signaling—providing new insight into neurotrophic factors in cancer(Man et al. 2023). Integrated Bioinformatics analysis revealed significant enrichment of the eight genes are enriched in key TNBC pathways—including (PI3K-Akt signaling, ECM–receptor interaction, and focal adhesion). This suggesting a functional coupling between tryptophan metabolic dysregulation and core oncogenic networks, collectively influencing tumor malignancy and the immune microenvironment. Together, these findings support the biological relevance of our prognostic model. In this study, ssGSEA was employed to reveal a distinct immune infiltration landscape in high-risk TNBC, characterized by significant enrichment of central memory CD8⁺ T cells, immature dendritic cells, and neutrophils. Correlation analysis demonstrated that PLAU expression was significantly positively correlated with the infiltration of these immune cell subsets, whereas NRTN showed a significant negative correlation with central memory CD8⁺ T cells and neutrophils. This landscape presents a paradoxical scenario—extensive immune cell infiltration without a corresponding survival benefit—suggesting an 'infiltrated but suppressed' phenotype in high-risk TNBC. We hypothesize that PLAU-mediated extracellular matrix remodeling and chemotaxis may act as key drivers of immune cell recruitment, while functional suppression arises from the synergistic interplay of multiple mechanisms. First, tryptophan metabolism via the kynurenine pathway promotes T cell exhaustion and regulatory T cell differentiation(Man et al. 2023), which can lead to disrupted T cell receptor signaling and upregulation of co-inhibitory molecules, driving CD8⁺ T cells into a state of "pseudo-exhaustion" characterized by retained migratory capacity but loss of effector function(Koushki et al. 2021, Anderson et al. 2017, Scarlett et al. 2012). Second, the accumulation of immature dendritic cells, which exhibit low antigen-presenting capacity, may further reinforce immunosuppression by inducing T cell anergy(Scarlett et al. 2012). In addition, tumor-associated neutrophils can directly suppress T cell function through the secretion of arginase-1 and reactive oxygen species(Liu et al. 2022). Collectively, these three layers form a multidimensional immunosuppressive network that accounts for the paradoxical association between extensive immune infiltration and poor prognosis. The negative correlation between NRTN and immune cell infiltration further suggests that this protective gene may exert its effects, at least in part, by constraining such dysfunctional immune recruitment, although the underlying mechanisms warrant experimental investigation. Beyond the immunosuppressive microenvironment features of high-risk TNBC highlighted earlier, molecular pathway dysregulation represents another critical driver of its aggressive phenotype and poor prognosis. GSVA of TNBC molecular heterogeneity revealed significant pathway activity differences between risk groups, offering insights into tumor biology and therapeutic opportunities. The high-risk group exhibited marked activation of pathways linked to substance transport, energy metabolism, and membrane function—including "ABC transporters," "N-glycan biosynthesis," "Type II diabetes," "Riboflavin metabolism," and "Endocytosis." These activated pathways collectively drive TNBC aggressiveness, therapy resistance, and poor prognosis. Specifically, ABC transporter upregulation in high-risk TNBC enhances drug efflux, reducing intracellular chemotherapeutic concentrations and driving multidrug resistance. Complementing the immunosuppressive features observed in high-risk TNBC, GSVA-based pathway enrichment analyses further delineate the molecular underpinnings of its aggressive phenotype. The coordinated activation of ABC transporters and endocytosis pathways in the high-risk group points to concurrently enhanced drug efflux and membrane trafficking(Chen et al. 2020, Egea et al. 2022). Consistently, this group showed significantly higher IC₅₀ values for most chemotherapeutic agents, suggesting a broadly refractory phenotype that may underlie chemoresistance in high-risk TNBC. Such a resistance barrier likely represents a critical clinical determinant of poor prognosis and may be potentially overcome by combining ABC transporter inhibitors with chemotherapy(Chien et al. 2025). Activation of N-glycan biosynthesis may synergize with PLAU-mediated extracellular matrix remodeling to facilitate tumor invasion and immune evasion. One potential mechanism involves altered glycosylation of adhesion molecules such as integrins, which can modulate cell–matrix interactions and immune recognition(Shi et al. 2020 , Xue et al. 2023b ). Moreover, glycosylated proteins derived from this pathway hold promise as diagnostic biomarkers or therapeutic targets(Ščupáková et al. 2021 ). Enrichment of the type 2 diabetes mellitus pathway aligns with the lipid metabolic reprogramming indicated by early high expression of FABP7 in the high-risk group, suggesting that concurrent dysregulation of glucose and lipid metabolism supports tumor proliferation and survival. Notably, exosome-mediated signaling under insulin-resistant conditions has been reported to promote epithelial–mesenchymal transition(Qiu et al. 2025 ), further linking metabolic disturbances to aggressive behavior. Conversely, suppression of cysteine and methionine metabolism, along with reduced keratan sulfate biosynthesis, suggests impaired glutathione synthesis and altered extracellular matrix sulfation. These changes may contribute to disease progression by increasing oxidative stress and promoting matrix remodeling(Park et al. 2022, Upadhyayula et al. 2023, Leiphrakpam et al. 2019), potentially creating a pro-tumorigenic microenvironment. Collectively, high-risk TNBC exhibits a synergistic pattern characterized by enhanced transport, metabolic reprogramming, and matrix dysregulation, which closely aligns with the multidimensional malignant features captured by the prognostic genes identified in this study. From a translational perspective, these findings highlight several candidate strategies for improving outcomes in high-risk patients, including combining ABC transporter inhibitors with chemotherapy to counter chemoresistance, targeting N-glycan biosynthesis pathways to potentially restore antitumor immunity, and modulating sulfur amino acid metabolism to disrupt metabolic adaptation. Our single-cell transcriptomic analysis provided unprecedented resolution into the TNBC tumor microenvironment (TME). Beyond confirming its immune-rich, stroma-poor nature, we delineated cell type-specific expression of eight core prognostic genes in T cells, B cells, macrophages, and stromal cells. Pseudotime analysis further revealed the dynamic expression of these eight prognostic genes during key cellular state transitions. Notably, FABP7 and PLAU displayed distinct temporal expression patterns during macrophage differentiation: FABP7 peaked early, while PLAU was markedly upregulated later, thereby suggesting stage-specific functional specialization. As a key lipid metabolism regulator, FABP7 facilitates fatty acid uptake and transport, thereby supporting tumor survival and proliferation. FABP7 may also influence energy homeostasis and the cell cycle via the PPAR-α signaling pathway(Kwong et al. 2020, Kwong et al. 2019, Miyazaki et al. 2025, Xu et al. 2020). Its early expression likely supplies lipids and energy to drive M2 polarization, thereby fostering an immunosuppressive microenvironment(Miyazaki et al. 2025, Xu et al. 2025). In contrast, PLAU—a central element of the plasminogen activation system—promotes tumor invasion, angiogenesis, and the ECM degradation(Shi et al. 2024 ). Late PLAU upregulation may indicate macrophage maturation, thereby enabling local invasion and metastasis via ECM breakdown(Zhu et al. 2021 , Wan et al. 2025). These temporal patterns outline a functional progression: early metabolic remodeling via FABP7 supports polarization, while late PLAU expression facilitates pro-invasive ECM degradation. Therapeutically, early FABP7 inhibition may disrupt metabolic adaptation and M2 macrophage polarization, while late PLAU targeting may suppress TNBC metastasis. Additionally, the sustained high expression of EIF4EBP1 and PLAU in stromal cells indicates that stromal cells are highly activated cancer-associated fibroblasts (CAFs) with enhanced protein synthesis and ECM secretion—consistent with reported fibroblast activation programs(Torrence et al. 2021) —thereby corroborating, at single-cell resolution, the active stromal remodeling in high-risk TNBC. Based on the cell-specific expression patterns of prognostic genes revealed by single-cell analysis, we further employed the GraphBAN model to screen for potential therapeutic compounds targeting these genes. A key translational finding of this study is the utilization of the GraphBAN model—a graph neural network (GNN)—for translating molecular discoveries into clinical therapy. This AI model systematically predicted interactions between ω-3 carboxylic acid and key prognostic components such as COL9A3. While a long-chain polyunsaturated fatty acid, ω-3 carboxylic acid, is known for its anti-inflammatory, antioxidant, and potential anti-tumor properties(Nicholls et al. 2020, Cikoš et al. 2021), its specific targets and mechanisms in TNBC remain unclear. Our model proposes a novel hypothesis: ω-3 carboxylic acid may exert anti-TNBC effects by binding specifically to COL9A3, a notion requiring experimental validation. COL9A3, a component of type IX collagen, is overexpressed in aggressive tumors and may promote malignancy by activating integrin-mediated FAK/Src and PI3K/Akt signaling, thereby inducing EMT and enhancing migration and invasion(Yang et al. 2024 , Wu et al. 2021b). Based on this, we speculate that ω-3 carboxylic acids may suppress the pro-tumor function of COL9A3 by interfering with its interaction with integrins, a mechanism that requires experimental validation. This insight suggests novel therapeutic avenues for TNBC via targeted ECM modulation. Notably, ω-3 carboxylic acid, an FDA-approved agent for hypertriglyceridemia, is increasingly studied in cancer for its anti-inflammatory and potential antitumor effects. Given the tryptophan metabolism-driven immunosuppressive microenvironment identified in our study, its immunomodulatory properties may further counterbalance this effect, offering a rationale for combination therapy. We therefore propose ω-3 carboxylic acid as a priority candidate for TNBC therapeutic development. Preclinical evaluation should confirm binding to COL9A3, assess monotherapy and combination therapy efficacy in animal models, and elucidate effects on the TME—notably immune cell infiltration and function. This AI-guided approach underscores the potential of computational prediction to accelerate translational oncology. This study established and validated a robust tryptophan metabolism-related prognostic model for TNBC via integrated multi-omics analysis, systematically elucidating its biological basis at the pathway, TME, and functional levels: it correlates with PI3K-Akt signaling activation and ECM remodeling, reflects an "infiltrated but suppressed" immune state (characterized by tryptophan metabolic dysregulation, T cell pseudo-exhaustion, and neutrophil-mediated immunosuppression), and captures enhanced ABC transporter-mediated drug efflux plus metabolic adaptations of riboflavin metabolism and type II diabetes-related pathways. Additionally, single-cell transcriptomics and pseudotime analysis mapped prognostic genes to specific cellular subsets and revealed dynamic expression of key genes (e.g., FABP7, PLAU) during macrophage differentiation (yielding temporal-spatial functional insights), while the GraphBAN AI model identified potential interactions between approved drugs (e.g., ω-3 carboxylic acid) and core targets (e.g., COL9A3) to propose novel therapeutic candidates. A key limitation is its reliance on public datasets, which requires future experimental validation. Future work will: (1) functionally validate core genes (e.g., EIF4EBP1, NRTN) in TNBC malignancy; (2) elucidate their mechanisms in tryptophan metabolism-mediated immunosuppression; and (3) preclinically evaluate predicted drugs, including ω-3 carboxylic acid, alone and combined with chemotherapy or immunotherapy. In summary, this study provides a practical prognostic tool for TNBC, deepens multi-omics understanding of tryptophan metabolic reprogramming in progression, and offers translatable leads for targeted and combination immunotherapy. 5 Conclusion Our integrated multi-omics study constructed and validated an eight-gene prognostic signature—centered on tryptophan metabolism (EIF4EBP1, NRTN, COL9A3, etc.)—for TNBC. The RSF model was identified as the optimal model, generating a risk score that independently stratified patient survival in both TCGA and GEO cohorts while outperforming conventional clinical predictors. Biologically, the signature reflects coordinated PI3K-Akt pathway activation and extracellular matrix (ECM) remodeling, and correlates with an immunosuppressive tumor microenvironment characterized by dysregulated tryptophan metabolism, T cell dysfunction, and neutrophil-mediated suppression. Single-cell analyses revealed differentiation-associated expression dynamics of FABP7 and PLAU in macrophages. GraphBAN AI modeling further predicted ω-3 carboxylic acid interaction with COL9A3, proposing a new therapeutic avenue. Although experimental confirmation is needed, this work delivers a robust prognostic tool, enhances multi-omics understanding of tryptophan rewiring in TNBC, and suggests actionable targets for combination therapy. Abbreviation scRNA-seq Single-Cell RNA Sequencing / Transcriptomics and RNA Sequencing GEO Gene Expression Omnibus QC Quality Control PCs Principal Components UMAP Uniform Manifold Approximation and Projection Declarations Data Availability Statement The datasets analysed in this study are available in The Cancer Genome Atlas (TCGA) repository (https://portal.gdc.cancer.gov/)，in Gene Expression Omnibus (GEO) database (https://www.ncbi. nlm.nih.gov/gds), including GSE135565 and GSE161529 Conflict of i nterest The authors declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Author Contributions Conceptualization: Youjun Wu; data curation: Xiaorong Pang; formal analysis: Feng Cen; investigation: Liang Xie, methodology: Youjun Wu; resources: Xianglan Mo; software: Xiang Feng; supervision: Xiaorong Pang; validation: Feng Cen; visualization: Xiang Feng; writing—original draft: Youjun Wu; writing—review and editing: Xianglan Mo. All authors read and approved the final manuscript. Funding This study was supported by the Guangxi Natural Science Foundation Project (grant number：2023GXNSFBA026013) and the Science and Technology Base and Talent Program in Guangxi (grant number：GuikeAD23026097). Acknowledgments We would like to express our sincere gratitude to all individuals and organizations who supported and assisted us throughout this research. Without your support, this study would not have been feasible. Consent for Publication Not applicable. Ethics approval and consent to patients Not applicable. References ANDERSON K. G. et al. (2017) Obstacles Posed by the Tumor Microenvironment to T cell Activity: A Case for Synergistic Therapies. Cancer Cell 31(3):311-325 ARAN D. et al. (2019) Reference-based analysis of lung single-cell sequencing reveals a transitional profibrotic macrophage. Nat Immunol 20(2):163-172 BRAY F. et al. (2024) Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 74(3):229-263 CAO J. et al. (2019) The single-cell transcriptional landscape of mammalian organogenesis. Nature 566(7745):496-502 CHANG C. H. et al. (2019) A novel orally available seleno-purine molecule suppresses triple-negative breast cancer cell proliferation and progression to metastasis by inducing cytostatic autophagy. Autophagy 15(8):1376-1390 CHAROENTONG P. et al. (2017) Pan-cancer Immunogenomic Analyses Reveal Genotype-Immunophenotype Relationships and Predictors of Response to Checkpoint Blockade. Cell Rep 18(1):248-262 CHEN H. & BOUTROS, P. C. (2011) VennDiagram: a package for the generation of highly-customizable Venn and Euler diagrams in R. BMC Bioinformatics 12:35 CHEN P. S. et al. (2020) Pathophysiological implications of hypoxia in human diseases. J Biomed Sci 27(1):63 CHENG Q. et al. (2021) Three hematologic/immune system-specific expressed genes are considered as the potential biomarkers for the diagnosis of early rheumatoid arthritis through bioinformatics analysis. J Transl Med 19(1):18 CHIEN P. Y. et al. (2025) 2',6'-dihydroxy-3',4'-dimethoxydihydrochalcone counteracts cancer multidrug resistance by impeding STAT3 activation and ABC transporter-mediated drug efflux. Biomed Pharmacother 188:118153 CIKOŠ A. M. et al. (2021) Bioprospecting of Coralline Red Alga Amphiroa rigida J.V. Lamouroux: Volatiles, Fatty Acids and Pigments. Molecules 26(3):520 CORREIA J. C. et al. (2021) Muscle-secreted neurturin couples myofiber oxidative metabolism and slow motor neuron identity. Cell Metab 33(11):2215-2230.e8 DOWLING C. M. et al. (2021) Multiple screening approaches reveal HDAC6 as a novel regulator of glycolytic metabolism in triple-negative breast cancer. Sci Adv 7(3): eabc4897 EGEA J. et al. (2022) Alkylating Agent-Induced Toxicity and Melatonin-Based Therapies. Front Pharmacol 13:873197 FANG Z. et al. (2022) Overactivation of hepatic mechanistic target of rapamycin kinase complex 1 (mTORC1) is associated with low transcriptional activity of transcription factor EB and lysosomal dysfunction in dairy cows with clinical ketosis. J Dairy Sci 105(5):4520-4533 GREENE L. I. et al. (2019) A Role for Tryptophan-2,3-dioxygenase in CD8 T-cell Suppression and Evidence of Tryptophan Catabolism in Breast Cancer Patient Plasma. Mol Cancer Res 17(1):131-139 GU Z. & HüBSCHMANN, D. (2022) Make Interactive Complex Heatmaps in R. Bioinformatics 38(5):1460-1462 GUO B. et al. (2020) Micropeptide CIP2A-BP encoded by LINC00665 inhibits triple-negative breast cancer progression. Embo j 39(1):e102190 GUSTAVSSON E. K. et al. (2022) ggtranscript: an R package for the visualization and interpretation of transcript isoforms using ggplot2. Bioinformatics 38(15):3844-3846 HADIPOUR H. et al. (2025) GraphBAN: An inductive graph-based approach for enhanced prediction of compound-protein interactions. Nat Commun 16(1):2541 HANAHAN D. (2022) Hallmarks of Cancer: New Dimensions. Cancer Discov 12(1):31-46 HäNZELMANN S. et al. (2013) GSVA: gene set variation analysis for microarray and RNA-seq data. BMC Bioinformatics 14:7 HEAGERTY P. J. et al. (2000) Time-dependent ROC curves for censored survival data and a diagnostic marker. Biometrics 56(2):337-344 JAHCHAN N. S. et al. (2019) Tuning the Tumor Myeloid Microenvironment to Fight Cancer. Front Immunol 10:1611 JI P. et al. (2022) In vivo multidimensional CRISPR screens identify Lgals2 as an immunotherapy target in triple-negative breast cancer. Sci Adv 8(26):eabl8247 JIN S. et al. (2024) A prognostic model for overall survival in recurrent glioma patients treated with bevacizumab-containing therapy. Discov Oncol 15(1):85 KOUSHKI K. et al. (2021) Role of myeloid-derived suppressor cells in viral respiratory infections; Hints for discovering therapeutic targets for COVID-19. Biomed Pharmacother 144:112346 KUO L. W. et al. (2024) Blocking Tryptophan Catabolism Reduces Triple-Negative Breast Cancer Invasive Capacity. Cancer Res Commun 4(10):2699-2713 KWONG S. C. et al. (2020) Fatty acid binding protein 7 mediates linoleic acid-induced cell death in triple negative breast cancer cells by modulating 13-HODE. Biochimie 179:23-31 KWONG S. C. et al. (2019) Metabolic role of fatty acid binding protein 7 in mediating triple-negative breast cancer cell death via PPAR-α signaling. J Lipid Res 60(11):1807-1817 LEI J. et al. (2023) Clinicopathological characteristics of pheochromocytoma/paraganglioma and screening of prognostic markers. J Surg Oncol 128(4):510-518 LEIPHRAKPAM P. D. et al. (2019) Role of keratan sulfate expression in human pancreatic cancer malignancy. Sci Rep 9(1):9665 LERCHER A. et al. (2020) Systemic Immunometabolism: Challenges and Opportunities. Immunity 53(3):496-509 LIN M. et al. (2025) Integrating multi-omics data of Triple-Negative Breast Cancer to explore the role of Kynurenine pathway and KYNU as a therapeutic target. Biochem Biophys Res Commun 756:151569 LIU P. et al. (2022) Pretreatment Systemic Immune-Inflammation Index Can Predict Response to Neoadjuvant Chemotherapy in Cervical Cancer at Stages IB2-IIB. Pathol Oncol Res 28:1610294 LIU P. et al. (2020) Potential Molecular Mechanisms of Plantain in the Treatment of Gout and Hyperuricemia Based on Network Pharmacology. Evid Based Complement Alternat Med 2020:30231273023127 LIU T. T. et al. (2021) Identification of CDK2-Related Immune Forecast Model and ceRNA in Lung Adenocarcinoma, a Pan-Cancer Analysis. Front Cell Dev Biol 9:682002 LOVE M. I. et al. (2014) Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol 15(12):550 MA P. et al. (2023) Bacterial droplet-based single-cell RNA-seq reveals antibiotic-associated heterogeneous cellular states. Cell 186(4):877-891.e14 MAN J. et al. (2023) TANGO1 interacts with NRTN to promote hepatocellular carcinoma progression by regulating the PI3K/AKT/mTOR signaling pathway. Biochem Pharmacol 213:115615 MIYAZAKI H. et al. (2025) FABP7 in Hepatic Macrophages Promotes Fibroblast Activation and CD4(+) T-Cell Migration by Regulating M2 Polarization During Liver Fibrosis. J Immunol Res 2025:6987981 NICHOLLS S. J. et al. (2020) Effect of High-Dose Omega-3 Fatty Acids vs Corn Oil on Major Adverse Cardiovascular Events in Patients at High Cardiovascular Risk: The STRENGTH Randomized Clinical Trial. Jama 324(22):2268-2280 PARK J. Y. et al. (2022) Dysbiotic change in gastric microbiome and its functional implication in gastric carcinogenesis. Sci Rep 12(1):4285 QIU Y. et al. (2025) Insulin Resistance Increases TNBC Aggressiveness and Brain Metastasis via Adipocyte-Derived Exosomes. Mol Cancer Res 23(6):567-578 RITCHIE M. E. et al. (2015) limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res 43(7):e47 ROGERS T. J. et al. (2019) Reversal of Triple-Negative Breast Cancer EMT by miR-200c Decreases Tryptophan Catabolism and a Program of Immunosuppression. Mol Cancer Res 17(1):30-41 SATIJA R. et al. (2015) Spatial reconstruction of single-cell gene expression data. Nat Biotechnol 33(5):495-502 SCARLETT U. K. et al. (2012) Ovarian cancer progression is controlled by phenotypic changes in dendritic cells. J Exp Med 209(3):495-506 ŠČUPáKOVá K. et al. (2021) Clinical importance of high-mannose, fucosylated, and complex N-glycans in breast cancer metastasis. JCI Insight 6(24): e146945 SHI K. et al. (2024) Pan-cancer analysis of PLAU indicates its potential prognostic value and correlation with neutrophil infiltration in BLCA. Biochim Biophys Acta Mol Basis Dis 1870(2):166965 SHI S. et al. (2020) Inhibition of MAN2A1 Enhances the Immune Response to Anti-PD-L1 in Human Tumors. Clin Cancer Res 26(22):5990-6002 SUGIAWAN Y. et al. (2023) Assessing the United Nations sustainable development goals from the inclusive wealth perspective. Sci Rep 13(1):1601 TAN Z. et al. (2025) Machine Learning and Experimental Validation Reveal MYH11 as a Novel Prognostic Biomarker and Therapeutic Target in Bladder Cancer. J Inflamm Res 18:8357-8387 TAN Z. et al. (2023) Comprehensive analysis of scRNA-Seq and bulk RNA-Seq reveals dynamic changes in the tumor immune microenvironment of bladder cancer and establishes a prognostic model. J Transl Med 21(1):223 TORRENCE M. E. et al. (2021) The mTORC1-mediated activation of ATF4 promotes protein and glutathione synthesis downstream of growth signals. Elife 10:e63326 UPADHYAYULA P. S. et al. (2023) Dietary restriction of cysteine and methionine sensitizes gliomas to ferroptosis and induces alterations in energetic metabolism. Nat Commun 14(1):1187 UTHAYOPAS K. et al. (2021) TSMDA: Target and symptom-based computational model for miRNA-disease-association prediction. Mol Ther Nucleic Acids 26:536-546 VEGHINI L. et al. (2024) Differential activity of MAPK signalling defines fibroblast subtypes in pancreatic cancer. Nat Commun 15(1):10534 WAN L. et al. (2025) Hypoxia-induced tumor cell-intrinsic PLAU activation drives immunotherapy resistance in collagenic lung adenocarcinoma. Int Immunopharmacol 162:115161 WANG Y. et al. (2020) LncRNA-encoded polypeptide ASRPS inhibits triple-negative breast cancer angiogenesis. J Exp Med http://doi.org/10.1084/jem.20190950 WU T. et al. (2021a) clusterProfiler 4.0: A universal enrichment tool for interpreting omics data. Innovation (Camb) 2(3):100141 WU W. et al. (2021b) Drivers and suppressors of triple-negative breast cancer. Proc Natl Acad Sci U S A 118(33): e2104162118 WU X. et al. (2024) Tumoral EIF4EBP1 regulates the crosstalk between tumor-associated macrophages and tumor cells in MRTK. Eur J Pharmacol 978:176787 XU D. et al. (2020) Upregulation of FABP7 inhibits acute kidney injury-induced TCMK-1 cell apoptosis via activating the PPAR gamma signalling pathway. Mol Omics 16(6):533-542 XU S. et al. (2025) FABP7-mediated lipid-laden macrophages drive the formation of pre-metastatic niche and liver metastasis. Int J Biol Sci 21(10):4388-4409 XUE C. et al. (2023a) Tryptophan metabolism in health and disease. Cell Metab 35(8):1304-1326 XUE Z. et al. (2023b) Investigation on acquired palbociclib resistance by LC-MS based multi-omics analysis. Front Mol Biosci 10:1116398 YAN J. et al. (2024) Molecular mechanisms and therapeutic significance of Tryptophan Metabolism and signaling in cancer. Mol Cancer 23(1):241 YANG X. et al. (2024) Potential regulation and prognostic model of colorectal cancer with extracellular matrix genes. Heliyon 10(16):e36164 YU L. et al. (2022) A Retrospective and Multicenter Study on COVID-19 in Inner Mongolia: Evaluating the Influence of Sampling Locations on Nucleic Acid Test and the Dynamics of Clinical and Prognostic Indexes. Front Med (Lausanne) 9:830484 ZHANG D. et al. (2022a) Novel insight on marker genes and pathogenic peripheral neutrophil subtypes in acute pancreatitis. Front Immunol 13:964622 ZHANG D. et al. (2022b) Facile synthesis of near-infrared responsive on-demand oxygen releasing nanoplatform for precise MRI-guided theranostics of hypoxia-induced tumor chemoresistance and metastasis in triple negative breast cancer. J Nanobiotechnology 20(1):104 ZHANG Z. et al. (2023) Mitochondrial energy metabolism correlates with an immunosuppressive tumor microenvironment and poor prognosis in esophageal squamous cell carcinoma. Comput Struct Biotechnol J 21:4118-4133 ZHOU F. et al. (2021) Photo-activated chemo-immunotherapy for metastatic cancer using a synergistic graphene nanosystem. Biomaterials 265:120421 ZHU T. et al. (2021) Identification of a Competing Endogenous RNA Network Related to Immune Signature in Lung Adenocarcinoma. Front Genet 12:665555 Additional Declarations No competing interests reported. Supplementary Files Supplementarymaterial.docx Supplementary Figure S1 The scRNA-seq data processing. (A) After QC of GSE161529; (B) The top 2000 HVGs; (C-D) PCA of GSE161529. Supplementary Table S1 Summary Table of DEGs. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9287896","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":616479079,"identity":"f8ce531b-90b4-4469-97d1-1f3a57379232","order_by":0,"name":"Youjun Wu","email":"","orcid":"","institution":"The First Affiliated Hospital of Jinan University","correspondingAuthor":false,"prefix":"","firstName":"Youjun","middleName":"","lastName":"Wu","suffix":""},{"id":616479080,"identity":"24b7d490-ea3e-43a8-b919-e4b1fb58f4fa","order_by":1,"name":"Xiaorong Pang","email":"","orcid":"","institution":"The People’s Hospital of Guangxi Zhuang Autonomous Region, Guangxi Academy of Medical Sciences","correspondingAuthor":false,"prefix":"","firstName":"Xiaorong","middleName":"","lastName":"Pang","suffix":""},{"id":616479081,"identity":"f870d7f6-1540-4da2-8523-e63d41daafe1","order_by":2,"name":"Feng Cen","email":"","orcid":"","institution":"The People’s Hospital of Guangxi Zhuang Autonomous Region, Guangxi Academy of Medical Sciences","correspondingAuthor":false,"prefix":"","firstName":"Feng","middleName":"","lastName":"Cen","suffix":""},{"id":616479082,"identity":"70e249e5-22a7-49ec-add7-85fbe1fa4dcd","order_by":3,"name":"Liang Xie","email":"","orcid":"","institution":"The People’s Hospital of Guangxi Zhuang Autonomous Region, Guangxi Academy of Medical Sciences","correspondingAuthor":false,"prefix":"","firstName":"Liang","middleName":"","lastName":"Xie","suffix":""},{"id":616479083,"identity":"94b6eb6e-aafa-46f6-9ab9-6508c6e6d308","order_by":4,"name":"Xiang Feng","email":"","orcid":"","institution":"The People’s Hospital of Guangxi Zhuang Autonomous Region, Guangxi Academy of Medical Sciences","correspondingAuthor":false,"prefix":"","firstName":"Xiang","middleName":"","lastName":"Feng","suffix":""},{"id":616479084,"identity":"f60dc74b-0ce6-40d1-a62c-1929a5cf794d","order_by":5,"name":"Xianglan Mo","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA0klEQVRIiWNgGAWjYJCCA0Asw8DAfPBBAilaeBgY2JINHpBiE1ALj5kkUVp023sMD/zcUcvDP7vng0RCzR0G/vZu/O4zO3Ms4WDvmeM8EnfObjBIOPaMQeLM2Q34tdxIPnCAt+0YD8ON3A0JiQ2HGQwkcglpSWw4+BeoRf5GzoMDRGpJPnCYt62Gx+BGDmMDcVqAfjks23aAx/BGmjFDwrHDPIT9crzH+OPbtjo5uRvJz3/+qDksx9/ei18LFByGs3iIUQ4CdcQqHAWjYBSMgpEIAL82UgCXC6shAAAAAElFTkSuQmCC","orcid":"","institution":"The People’s Hospital of Guangxi Zhuang Autonomous Region, Guangxi Academy of Medical Sciences","correspondingAuthor":true,"prefix":"","firstName":"Xianglan","middleName":"","lastName":"Mo","suffix":""}],"badges":[],"createdAt":"2026-04-01 07:09:28","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9287896/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9287896/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":106415595,"identity":"e4dc8dda-6062-4a06-8ca2-403c7b47252f","added_by":"auto","created_at":"2026-04-08 10:35:31","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":29951780,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDiscovery of 48 candidate genes and investigation into their biological roles.\u003c/strong\u003e (A-B) Volcano plot and heatmap of DEGs between TNBC samples and control samples; (C) Venn diagram of DEGs and MRGs; (D) The forest plot of the univariate Cox regression analysis; (E) The results of PH assumption; (F-G) GO and KEGG analyses of candidate genes; (H) PPI network of candidate genes.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-9287896/v1/8a1d8baa284ad17cb4fb6c04.png"},{"id":106415417,"identity":"ce91b925-6d03-444c-9453-6d5c590dd47d","added_by":"auto","created_at":"2026-04-08 10:34:06","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":6588019,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eConstruction and interpretation of the optimal model.\u003c/strong\u003e (A-C) The results of SHAP analysis; (D) Waterfall chart for each single sample.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-9287896/v1/cf15918329a7d94551656278.png"},{"id":106415420,"identity":"3257bce4-745d-4c16-a3ba-4e2054bdaa2a","added_by":"auto","created_at":"2026-04-08 10:34:06","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":10452923,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDevelopment and validation of the prognostic model.\u003c/strong\u003e (A) K-M curves of TCGA-TNBC and GSE135565; (B) Scatter plots of survival status shown in TCGA-TNBC and GSE135565; (C) ROC of 1-, 2-, 3-year in TCGA-TNBC and GSE135565.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-9287896/v1/9e306b709d4edf5f5d57c05c.png"},{"id":106415514,"identity":"706b17ee-602e-4f2f-8d96-40028a250528","added_by":"auto","created_at":"2026-04-08 10:34:43","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":8503914,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eIndependent prognostic analysis.\u003c/strong\u003e (A-D) Independent prognostic factors screened by univariate and multivariate Cox regression analyses and PH assumption test.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-9287896/v1/56c1c2d33463e8e496bb1ec4.png"},{"id":106415418,"identity":"bb574a5c-ce7f-4814-a7a3-d3783b3cf892","added_by":"auto","created_at":"2026-04-08 10:34:06","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":22158266,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTumor microenvironment and functional enrichment analyses.\u003c/strong\u003e (A) Immune cell infiltration heatmap of HRG and LRG; (B) Box plot of differences in infiltration levels of distinct immune cell types; (C) Correlation heatmap of prognostic genes and differentially infiltrated immune cells; (D) GSVA of HRG and LRG; (E) GSEA between HRG and LRG.\u003c/p\u003e","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-9287896/v1/2bbabe39421581c8549225da.png"},{"id":106415419,"identity":"b5d5f317-a6ae-4521-8dd6-c0d9aa03f1cf","added_by":"auto","created_at":"2026-04-08 10:34:06","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":9493672,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSomatic mutational landscape and prognostic value of TMB.\u003c/strong\u003e (A) Waterfall plots of gene mutation profiles; (B) Box plot of TMB variation between HRG and LRG; (C) K-M curves of high- and low-TMB groups; (D) K-M curves of high risk/high TMB, high risk/low TMB, low risk/high TMB, and low risk/low TMB groups.\u003c/p\u003e","description":"","filename":"Figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-9287896/v1/b1b145ad7e4b2fe87ddb19ed.png"},{"id":106415749,"identity":"01c9dea8-f367-4e51-8835-dd26788a7621","added_by":"auto","created_at":"2026-04-08 10:39:00","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":2871037,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAI-predicted therapeutic compounds and drug sensitivity.\u003c/strong\u003e (A) Network diagram of four prognostic genes and two compounds; (B) Drug sensitivity analysis between risk groups. *p \u0026lt; 0.05, **p \u0026lt; 0.01, ***p \u0026lt; 0.001.\u003c/p\u003e","description":"","filename":"Figure7.png","url":"https://assets-eu.researchsquare.com/files/rs-9287896/v1/8ae3859dd63ecc51a1871eb8.png"},{"id":106724009,"identity":"6259d56e-bfbf-4e43-8aeb-90bfb08a182a","added_by":"auto","created_at":"2026-04-12 18:23:48","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":15204807,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eIdentification of key cells.\u003c/strong\u003e (A) UMAP plot of 53 cell clusters; (B) Annotation of 8 distinct cell types based on marker gene expression; (C) Distribution proportions of different annotated cell types in the TNBC and control groups; (D) Expression levels of prognostic genes in different cell types; (E) Differences in the expression of prognostic genes across different cell types.\u003c/p\u003e","description":"","filename":"Figure8.png","url":"https://assets-eu.researchsquare.com/files/rs-9287896/v1/4c43e23923073d131e1fe9cc.png"},{"id":106415449,"identity":"327617ed-4da3-4f62-a545-0280c04af4e5","added_by":"auto","created_at":"2026-04-08 10:34:20","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":26179919,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePseudo-temporal trajectory analysis.\u003c/strong\u003e (A-B) Pseudo-temporal trajectory analysis of key cells; (C-E) Expression changes of prognostic genes observed during key cells differentiation.\u003c/p\u003e","description":"","filename":"Figure9.png","url":"https://assets-eu.researchsquare.com/files/rs-9287896/v1/e713f9401b95e7ef7043c739.png"},{"id":106415411,"identity":"02c4b0de-8aa5-424d-9caa-0e423dfa4c2b","added_by":"auto","created_at":"2026-04-08 10:34:02","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":22643932,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTF activity and metabolic pathway activity analyses.\u003c/strong\u003e (A) TF activity analysis of key cells; (B) Metabolic pathway activity analysis of key cells.\u003c/p\u003e","description":"","filename":"Figure10.png","url":"https://assets-eu.researchsquare.com/files/rs-9287896/v1/8941bae0c8ebc55490ee2eef.png"},{"id":106415516,"identity":"e669e867-6d3f-41d4-84ce-411a8c08aadd","added_by":"auto","created_at":"2026-04-08 10:34:43","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":18571256,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eIntercellular communication networks in the TNBC microenvironment.\u003c/strong\u003e (A) Network plots showing the number of intercellular interactions among distinct cell types in TNBC (left) and control (right) samples. (B) Network plots displaying the interaction strength among cell types in TNBC (left) and control (right) samples. (C) Bubble plots illustrating ligand-receptor signaling probabilities when B cells, T cells, macrophages, and stromal cells were considered as signal sources in TNBC (left) and control (right) samples. (D) Bubble plots depicting ligand-receptor signaling probabilities when B cells, T cells, macrophages, and stromal cells were considered as target cells in TNBC (left) and control (right) samples.\u003c/p\u003e","description":"","filename":"Figure11.png","url":"https://assets-eu.researchsquare.com/files/rs-9287896/v1/549c2fd40678ecc50c233182.png"},{"id":106092511,"identity":"e1ccea0a-acac-4608-b4a7-86dc6b5fa7ec","added_by":"auto","created_at":"2026-04-03 11:20:20","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1312731,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9287896/v1/cf5f971b-8d47-462d-a6c8-6297337806e0.pdf"},{"id":106415596,"identity":"d7ad0ea9-b393-4352-af7a-45d5259d73c9","added_by":"auto","created_at":"2026-04-08 10:35:31","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1393656,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSupplementary Figure S1\u003c/strong\u003e The scRNA-seq data processing. (A) After QC of GSE161529; (B) The top 2000 HVGs; (C-D) PCA of GSE161529.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSupplementary Table S1\u003c/strong\u003e Summary Table of DEGs.\u003c/p\u003e","description":"","filename":"Supplementarymaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-9287896/v1/d6e6353eb253c041a52b517f.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Integrated bulk and single-cell transcriptomic analysis reveals a tryptophan metabolism-driven prognostic signature and therapeutic landscape in triple- negative breast cancer","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eBreast cancer, the second most common cancer worldwide after lung cancer in terms of incidence, is the leading cause of cancer mortality(Bray et al. \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Triple-negative breast cancer (TNBC) is defined by the absence of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2)(Dowling et al. 2021). This unique molecular profile renders it refractory to endocrine therapy and anti-HER2 targeted agents(Wang et al. \u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e2020\u003c/span\u003e), which consequently establishes TNBC as the breast cancer subtype with the most aggressive clinical behavior and poorest prognosis. TNBC comprises 10\u0026ndash;20% of all breast cancers and is distinguished by a set of clinical features, including increased aggressiveness, elevated recurrence risk, and a proclivity for visceral dissemination(Zhou et al. \u003cspan citationid=\"CR74\" class=\"CitationRef\"\u003e2021\u003c/span\u003e, Zhang et al. 2022b). The therapeutic landscape for TNBC remains dominated by cytostatic chemotherapy, yet the benefits of this treatment have reached a plateau in improving patient outcomes(Chang et al. 2019). Although emerging modalities, notably immune checkpoint inhibitors, offer renewed hope for a subset of patients, their clinical application is hampered by limited response rates and a critical lack of predictive biomarkers to reliably identify beneficiaries(Ji et al. 2022). Consequently, deciphering the fundamental molecular drivers of TNBC to develop reliable prognostic tools and identify new therapeutic vulnerabilities is an urgent and critical research imperative.\u003c/p\u003e \u003cp\u003eMetabolic reprogramming, a core cancer hallmark proposed by Hanahan (Hanahan 2022) supports tumor growth and fosters an immunosuppressive microenvironment. The tryptophan-kynurenine pathway is a critical mediator of this metabolic reprogramming process(Yan et al. 2024, Xue et al. 2023a). In TNBC, up to 95% of tryptophan is metabolized via this route(Lin et al. 2025). The enzymes indoleamine 2,3-dioxygenase 1 (IDO1) and tryptophan 2,3-dioxygenase 2 (TDO2) produce kynurenines, which impair T-cell function through two distinct mechanisms: by causing local tryptophan depletion and, via activation of the aryl hydrocarbon receptor (AhR), promoting regulatory T cell (Treg) differentiation while inhibiting effector T cells, thereby enabling tumor immune evasion(Kuo et al. 2024, Rogers et al. 2019, Greene et al. 2019). The clinical failure of IDO1 inhibitors reveals metabolic network complexity and redundancy, which demands a systems-level understanding of the regulation of tryptophan metabolism and inter-pathway crosstalk(Lercher et al. 2020, Jahchan et al. 2019).\u003c/p\u003e \u003cp\u003eTraditional prognostic models based on bulk RNA sequencing, while informative of overall gene expression, are limited by their inability to dissect intratumoral cellular complexity and interactions(Tan et al. \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Single-cell RNA sequencing (scRNA-seq) overcomes this by providing high-resolution maps of the tumor microenvironment and tracing key events to specific cell types(Ma et al. 2023). However, scRNA-seq data alone is often insufficient for building robust clinical prognostic models. Thus, a combined approach\u0026mdash;leveraging the prognostic power of bulk RNA-seq data and the cellular resolution of single-cell RNA-seq (scRNA-seq) data\u0026mdash;represents a more powerful research strategy.\u003c/p\u003e \u003cp\u003eThe \"black-box\" nature of machine learning prognostic models impedes their clinical adoption. Explainable AI (XAI) approaches like SHapley Additive exPlanations (SHAP) mitigate this by quantifying feature-specific contributions, thus preserving predictive power while enabling biological interpretation(Uthayopas et al. 2021, Yu et al. 2022). In a parallel challenge within translational medicine, the inefficient conversion of basic research findings into clinical therapies remains a major obstacle. AI-based drug discovery models, including Graph Neural Networks (GNNs) like GraphBAN, are addressing this by systematically screening for target-interacting compounds, accelerating both drug repurposing and novel drug development(Hadipour et al. 2025).\u003c/p\u003e \u003cp\u003eThis study employs an integrated multi-omics and computational strategy to define the prognostic value of tryptophan metabolism in TNBC. Our methodology involved constructing a prognostic model based on bulk RNA-seq transcriptomic data, followed by SHAP-based analysis to elucidate the contribution of key genes to the model's predictions. Furthermore, we will leverage artificial intelligence-based drug prediction platforms to identify potential lead compounds targeting this pathway. Finally, scRNA-seq data was utilized to dynamically resolve the cell-type-specific mechanisms underlying these associations between tryptophan metabolism and prognosis at the microscopic level. The overarching goal of this research is to provide novel therapeutic targets and a conceptual framework to advance precision medicine for TNBC.\u003c/p\u003e"},{"header":"2 Materials and methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Data source\u003c/h2\u003e \u003cp\u003eOn July 21, 2024, we retrieved bulk RNA-sequencing data along with clinical characteristics for 100 TNBC samples and 113 normal breast tissue specimens from The Cancer Genome Atlas (TCGA) repository (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://portal.gdc.cancer.gov/\u003c/span\u003e\u003cspan address=\"https://portal.gdc.cancer.gov/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), designated as TCGA-TNBC cohort. For external validation purposes, we obtained the GSE135565 dataset (GPL570 platform) from the Gene Expression Omnibus (GEO) database (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.ncbi.nlm.nih.gov/geo/\u003c/span\u003e\u003cspan address=\"https://www.ncbi.nlm.nih.gov/geo/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), which contained 83 TNBC cases with corresponding survival outcomes. Additionally, scRNA-seq data (GSE161529, GPL18573 platform) encompassing 4 TNBC and 13 normal breast tissue specimens were acquired from GEO. Genes associated with tryptophan metabolism (TMRGs) were compiled from dual database sources on July 21. Initially, we searched GeneCards (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.genecards.org\u003c/span\u003e\u003cspan address=\"https://www.genecards.org\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) using \"tryptophan metabolism\" as the query term, yielding 4,951 protein-coding genes. From this pool, we retained 3,113 genes demonstrating a relevance score\u0026thinsp;\u0026ge;\u0026thinsp;10. Subsequently, we extracted 51 genes from the Molecular Signatures Database (MSigDB) (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.gsea-msigdb.org\u003c/span\u003e\u003cspan address=\"https://www.gsea-msigdb.org\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) across three gene collections: \"KEGG_TRYPTOPHAN_METABOLISM\", \"REACTOME_TRYPTOPHAN_CATABOLISM\", and \"WP_TRYPTOPHAN_METABOLISM\". Following consolidation and duplicate removal, we established a comprehensive TMRG collection comprising 3,116 genes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Identification of candidate genes\u003c/h2\u003e \u003cp\u003eUsing the DESeq2 package (v 1.40.2)(Love et al. 2014), we performed differential expression profiling between TNBC and control specimens within the TCGA-TNBC cohort. Genes exhibiting adjusted p-values (adj.p.val)\u0026thinsp;\u0026lt;\u0026thinsp;0.05 combined with absolute log\u003csub\u003e2\u003c/sub\u003e fold change (|log\u003csub\u003e2\u003c/sub\u003eFC|)\u0026thinsp;\u0026gt;\u0026thinsp;1 were classified as differentially expressed genes (DEGs). Visualization of DEGs was accomplished through volcano plots generated by ggplot2 package (v 3.4.1) (Gustavsson et al. 2022) and heatmaps created using ComplexHeatmap package (v 2.14.0)(Gu and H\u0026uuml;bschmann \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). The top 10 upregulated and downregulated genes, ordered by descending |log\u003csub\u003e2\u003c/sub\u003eFC|, were annotated in volcano plots, with their expression profiles further illustrated in heatmaps. We identified differentially expressed TMRGs (DE-TMRGs) through intersection of DEGs with the TMRG collection using VennDiagram package (v 1.7.1)(Chen and Boutros \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2011\u003c/span\u003e). To pinpoint DE-TMRGs significantly correlated with overall survival (OS), we employed univariate Cox proportional hazards regression via the survival package (v 3.7.0) (Lei et al. 2023)(hazard ratio (HR)\u0026thinsp;\u0026ne;\u0026thinsp;1, p\u0026thinsp;\u0026lt;\u0026thinsp;0.05). The proportional hazards (PH) assumption was validated for these significant genes, with those demonstrating PH test p-values\u0026thinsp;\u0026gt;\u0026thinsp;0.05 being selected as candidate genes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Functional enrichment and protein-protein interaction (PPI) analysis of candidate genes\u003c/h2\u003e \u003cp\u003eTo decipher the biological roles of selected candidate genes, we conducted Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment evaluations using clusterProfiler package (v 4.2.2)(Wu et al. \u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e2021a\u003c/span\u003e). (adj.p.val\u0026thinsp;\u0026lt;\u0026thinsp;0.05) GO analysis encompassed three ontological dimensions: biological processes (BPs), cellular components (CCs), and molecular functions (MFs). The top 5 pathways ranked by adj.p.val from each GO domain and KEGG pathways were visualized via ggplot2 package (v 3.4.1). To explore protein-level interactions among candidate genes, we constructed a protein-protein interaction (PPI) network utilizing the Search Tool for the Retrieval of Interacting Genes (STRING) database (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://cn.string-db.org/\u003c/span\u003e\u003cspan address=\"https://cn.string-db.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) with a minimum interaction score threshold of 0.15.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Machine learning model construction and SHapley additive exPlanations (SHAP) analysis\u003c/h2\u003e \u003cp\u003eWe developed nine distinct machine learning algorithms using candidate gene data and OS information from TCGA-TNBC: Cox proportional hazards (coxph), survival support vector machine (survivalsvm), stepwise Cox regression (Stepwise), elastic net, Ridge regression, Lasso regression, random survival forest (RandomForestSRC), partial least squares Cox regression (plsRcox), and extreme gradient boosting (xgboost). Algorithm performance was assessed through 5-fold cross-validation procedures. Performance indicators included concordance index (C-index), time-dependent area under the curve (AUC) at 1, 2, and 3 years, and Brier scores. The algorithm demonstrating the highest C-index alongside AUC values exceeding 0.6 was designated as optimal. We implemented SHAP analysis using shapviz package (v 0.9.7) (Sugiawan et al. \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) to interpret the optimal algorithm's predictions, generating multiple visualization formats including bar plots, beeswarm plots, dependence plots, waterfall plots, and force plots.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Prognostic model construction and validation\u003c/h2\u003e \u003cp\u003eWe constructed a random survival forest (RSF) algorithm using randomForestSRC package (v 3.3.1) (Jin et al. 2024)based on candidate genes from the TCGA cohort. Genes demonstrating importance scores\u0026thinsp;\u0026gt;\u0026thinsp;0.025 were designated as final prognostic markers. The algorithm's predictive output served as the risk score. Patient stratification into high-risk group (HRG) and low-risk group (LRG) was achieved using the optimal cut-point determined by the surv_cutpoint function from survminer package (v 0.4.9) (Liu et al. 2021)with minprop\u0026thinsp;=\u0026thinsp;0.4. Kaplan-Meier (K-M) survival curves with log-rank tests evaluated survival disparities between groups (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Furthermore, the expression heatmap of prognostic markers, risk score distribution, and patient survival status across HRG and LRG were comprehensively visualized using ggplot2 package (v 3.4.1). Predictive accuracy of the risk score was determined by calculating the area under the ROC curve (AUC) for 1, 2, and 3-year OS using survivalROC package (v 1.0.3.1)(Heagerty et al. 2000). This entire methodology was validated in the independent GSE135565 cohort.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Independent prognostic analysis\u003c/h2\u003e \u003cp\u003eWithin TNBC specimens with available survival data from TCGA-TNBC, we performed univariate and multivariate Cox regression analyses using survival package (v 3.7.0) to evaluate the independence of risk scores and clinical parameters (age, tumor stage, etc.). Variables exhibiting HR\u0026thinsp;\u0026ne;\u0026thinsp;1, p\u0026thinsp;\u0026lt;\u0026thinsp;0.2, and satisfying the PH assumption (p\u0026thinsp;\u0026gt;\u0026thinsp;0.05) were deemed independent prognostic factors.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e\u003cb\u003e2.7 Immune infiltration, gene set variation analysis (GSVA)\u003c/b\u003e, \u003cb\u003eand gene set enrichment analysis (GSEA)\u003c/b\u003e\u003c/h2\u003e \u003cp\u003eWe utilized GSVA package (v 1.50.0) (H\u0026auml;nzelmann et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2013\u003c/span\u003e) with the single-sample gene set enrichment analysis (ssGSEA) algorithm to quantify relative abundances of 28 immune cell subsets (Charoentong et al. 2017) within each TNBC specimen from TCGA-TNBC dataset. Disparities in immune cell infiltration between risk categories were evaluated using Wilcoxon rank-sum test (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Correlations between prognostic gene expression/risk scores and significantly altered immune cells were computed using ggpubr package (v 0.6.0) (Cheng et al. \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2021\u003c/span\u003e)(|correlation coefficient (cor)| \u0026gt; 0.3, p\u0026thinsp;\u0026lt;\u0026thinsp;0.05).\u003c/p\u003e \u003cp\u003eTo investigate underlying biological pathways exhibiting differential activation between HRG and LRG, we executed GSVA using GSVA package (v 1.42.0)(H\u0026auml;nzelmann et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2013\u003c/span\u003e). The reference gene collection employed was \"c2.cp.kegg.v2023.1.Hs.symbols.gmt\", a curated compilation of KEGG database pathways, obtained from MSigDB (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://www.gsea-msigdb.org/gsea/msigdb/index.jsp\u003c/span\u003e\u003cspan address=\"http://www.gsea-msigdb.org/gsea/msigdb/index.jsp\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). The ssGSEA algorithm computed enrichment scores for each pathway across every TNBC specimen in the TCGA-TNBC cohort. Subsequently, differential pathway activity between HRG and LRG was determined using limma package (v 3.54.0) (Ritchie et al. 2015)(|t| \u0026gt; 2, p\u0026thinsp;\u0026lt;\u0026thinsp;0.05).\u003c/p\u003e \u003cp\u003eTo further explore the biological pathway differences between HRG and LRG, GSEA was performed using the \"c2.cp.kegg.v2023.1.Hs.symbols.gmt\" gene set from the MSigDB. Differential expression analysis between risk groups was conducted using the DESeq2 package (v 1.40.2), and genes were ranked by log\u003csub\u003e2\u003c/sub\u003e FC. GSEA was implemented using the GSEA function in the clusterProfiler package (v 4.2.2), with significance thresholds set at p\u0026thinsp;\u0026lt;\u0026thinsp;0.05 and |normalized enrichment score (NES)| \u0026gt; 1.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e2.8 Tumor mutation burden (TMB) analysis\u003c/h2\u003e \u003cp\u003eSomatic mutation profiles for TNBC specimens from the TCGA cohort were retrieved and processed via TCGAmutations package (v 0.3.0)(Tan et al. \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Mutation data were analyzed and visualized using maftools package (v 2.16.0)(Zhang et al. \u003cspan citationid=\"CR73\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). TMB was computed as the total mutation count per megabase. TMB differences between risk categories were assessed using Wilcoxon test. TNBC patients across the entire cohort were stratified into high-TMB and low-TMB categories based on median TMB values. K-M survival curves were subsequently generated for these categories, with statistical significance of OS differences evaluated using log-rank test (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e2.9 AI-based drug prediction\u003c/h2\u003e \u003cp\u003eTo translate prognostic genes into therapeutic opportunities, we aimed to identify small-molecule compounds capable of binding proteins encoded by prognostic genes. The objective centered on computational screening for high-affinity interactions, establishing rationale for drug repurposing or novel therapeutic development targeting the identified prognostic pathway in TNBC. We employed the Graph-Based Attention Network (GraphBAN) algorithm to predict compound-protein interactions (CPIs). SMILES notations of compounds were sourced from ChEMBL database (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.ebi.ac.uk/chembl/\u003c/span\u003e\u003cspan address=\"https://www.ebi.ac.uk/chembl/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), while protein sequences were retrieved from UniProt (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.uniprot.org/\u003c/span\u003e\u003cspan address=\"https://www.uniprot.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). Molecular graph features of compounds were extracted using ChemBERTa and Graph Convolutional Networks (GCN). Protein embedding features were generated through 1D-CNN and Evolutionary Scale Modeling (ESM). These features were fed into the GraphBAN algorithm, which utilizes bidirectional attention mechanisms to predict interaction probabilities. Compound-protein pairs demonstrating interaction probabilities\u0026thinsp;\u0026gt;\u0026thinsp;0.8 were considered high-affinity candidates. Network visualization was accomplished using Cytoscape (v 3.10.2)(Liu et al. 2020).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e2.10 Drug sensitivity analysis\u003c/h2\u003e \u003cp\u003eTo assess the association between the prognostic risk model and chemotherapeutic response, drug sensitivity data were obtained from the Cancer Therapeutics Response Portal (CTRP) and the Genomics of Drug Sensitivity in Cancer (GDSC) database (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.cancerrxgene.org/\u003c/span\u003e\u003cspan address=\"https://www.cancerrxgene.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). The half-maximal inhibitory concentration (IC\u003csub\u003e50\u003c/sub\u003e) of 138 common chemotherapeutic and molecular targeted agents was estimated using the pRRophetic package (v 0.5), which employs a ridge regression model to predict drug response based on pre-treatment tumor gene expression profiles. IC\u003csub\u003e50\u003c/sub\u003e values were compared between HRG and LRG using the Wilcoxon test (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e2.11 scRNA-seq analysis\u003c/h2\u003e \u003cp\u003eWe employed Seurat package (v 5.1.0) (Satija et al. 2015) for scRNA-seq data processing. Quality control involved filtering low-quality cells (nFeature_RNA\u0026thinsp;\u0026lt;\u0026thinsp;200 or \u0026gt;\u0026thinsp;8000, mitochondrial gene percentage\u0026thinsp;\u0026gt;\u0026thinsp;10%). Data normalization was performed using NormalizeData function. The top 3000 highly variable genes (HVGs) were identified via FindVariableFeatures function. Principal component analysis (PCA) was executed, with optimal principal component (PC) numbers selected for downstream analysis based on elbow plots and JackStraw tests. Cell clustering was performed using FindNeighbors and FindClusters functions (resolution\u0026thinsp;=\u0026thinsp;2) and visualized through Uniform Manifold Approximation and Projection (UMAP). Cell type annotation was accomplished using SingleR package (v 2.2.0) (Aran et al. 2019)and manually refined with CellMarker database (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://ngdc.cncb.ac.cn/cellmarker/\u003c/span\u003e\u003cspan address=\"https://ngdc.cncb.ac.cn/cellmarker/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). Differential abundance of cell types between TNBC and control specimens was evaluated using Wilcoxon test (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Key cell types were defined as those exhibiting differential abundance and significant differential expression of prognostic genes (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05, Wilcoxon test).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e2.12 Pseudotime, cell-cell communication, transcription factor (TF), and metabolic activity analysis\u003c/h2\u003e \u003cp\u003ePseudotime trajectory inference was conducted using Monocle2 package (v 2.28.0)(Cao et al. 2019). Genes exhibiting dynamic expression patterns along pseudotime were identified and visualized. Transcription factor (TF) activity was inferred using decoupleR package (v 2.6.0)(Veghini et al. 2024). Metabolic pathway activity was quantified using scMetabolism package (v 0.2.1)(Zhang et al. 2022a). Dot plots visualized metabolic activity patterns across key cell types.\u003c/p\u003e \u003cp\u003eTo investigate intercellular communication networks within the tumor microenvironment, cell-cell interaction analysis was performed using the CellChat package (v 1.6.1) on the scRNA-seq dataset. This algorithm quantitatively infers and analyzes intercellular communication networks by integrating ligand-receptor interaction databases with gene expression data. The number and strength of intercellular interactions were calculated for each cell type, and communication probabilities were estimated using a statistical model. Signaling networks were visualized for both TNBC and control samples.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e2.13 Statistical analysis\u003c/h2\u003e \u003cp\u003eAll statistical analyses were conducted using R software (v 4.2.2). Unless stated otherwise, Wilcoxon test was employed to assess between-group differences with significance threshold set at p\u0026thinsp;\u0026lt;\u0026thinsp;0.05.\u003c/p\u003e \u003c/div\u003e"},{"header":"3 Results","content":"\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Identification and functional enrichment of tryptophan metabolism-related candidate genes\u003c/h2\u003e \u003cp\u003eA total of 6,136 DEGs were identified between TNBC and control tissues, including 3,656 upregulated and 2,480 downregulated genes (adj.p.val\u0026thinsp;\u0026lt;\u0026thinsp;0.05, |log\u003csub\u003e2\u003c/sub\u003eFC| \u0026gt; 1) (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA-B, \u003cb\u003eSupplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e\u003c/b\u003e). Among these, 1,058 overlapped with TMRGs and were designated DE-TMRGs (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eC). Univariate Cox regression analysis identified 54 DE-TMRGs significantly associated with OS (HR\u0026thinsp;\u0026ne;\u0026thinsp;1, p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eD). PH assumption testing retained 49 genes (PH test p\u0026thinsp;\u0026gt;\u0026thinsp;0.05) (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eE). After removing one gene (CYP11B1) with an exceptionally large HR (\u0026gt;\u0026thinsp;10,000), 48 genes were finalized as candidate genes for subsequent analysis.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eGO and KEGG enrichment analyses were performed to elucidate the potential biological functions and signaling pathways associated with the 48 candidate genes. The results indicated that these genes were significantly enriched in BPs critical for tumor progression, including ''response to estradiol'', ''stem cell development'', and ''response to oxygen levels'' (adj.p.val\u0026thinsp;\u0026lt;\u0026thinsp;0.05) (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eF). This enrichment pattern suggests that the candidate genes may orchestrate a pro-tumorigenic program by modulating hormone sensitivity, maintaining cancer stemness, and adapting to hypoxic stress within the tumor microenvironment (TME). In terms of MFs and CCs, the genes were predominantly associated with ''extracellular matrix (ECM) structural constituent'' and ''collagen-containing ECM'', highlighting their potential role in remodeling the ECM to facilitate tumor invasion, metastasis, and cell-matrix communication. KEGG pathway analysis further corroborated these findings, revealing significant enrichment in key oncogenic pathways, most notably the ''PI3K-Akt signaling pathway'' and ''ECM-receptor interaction'' (adj.p.val\u0026thinsp;\u0026lt;\u0026thinsp;0.05) (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eG). The enrichment of the PI3K-Akt pathway, a well-established driver of cell survival, proliferation, and metabolism in TNBC, implies that our candidate gene signature may contribute to tumor aggressiveness and therapy resistance through this axis. Concurrently, the enrichment in ECM-receptor interaction underscores a mechanistic link to enhanced cell adhesion, migration, and activation of integrin-mediated survival signals.\u003c/p\u003e \u003cp\u003eTo validate these functional associations at the protein level, a PPI network was constructed. The network, comprising 47 nodes and 254 edges, exhibited robust biological connectivity among the candidate genes (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eH). This highly interconnected network suggests that these proteins do not function in isolation but rather as coordinated modules or complexes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Construction and interpretation of the optimal model\u003c/h2\u003e \u003cp\u003eNine machine learning models were trained and evaluated via 5-fold cross-validation. The RSF model demonstrated the highest C-index and maintained AUCs\u0026thinsp;\u0026gt;\u0026thinsp;0.6 across 1, 2, and 3 years, establishing it as the optimal model for prognosis prediction (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). SHAP analysis was applied to interpret the RSF model. The bar plot ranked the importance of the candidate genes, with NRTN and EIF4EBP1 showing high impacts (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA). A beeswarm plot summarized the distribution of SHAP values for each gene across all individual patients in the TCGA cohort (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB). The plot revealed that high expression of EIF4EBP1 (orange points clustered on the positive SHAP value side) was consistently associated with an increased risk prediction, whereas high expression of NRTN (orange points clustered on the negative SHAP value side) was associated with a decreased risk prediction. Notably, NRTN showed a wide dispersion of SHAP values, indicating its effect size varied considerably between patients. Dependence plots for the 4 most important genes demonstrate that increasing expression of EIF4EBP1 increased the predicted risk, while increasing expression of NRTN, FABP7, and COL9A3 decreased it (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eC). This reinforced their roles as potential risk and protective factors, respectively.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSelecting the optimal machine learning model\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eC_index\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAUC_1yr\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAUC_2yr\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAUC_3yr\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eBrier_1yr\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eBrier_2yr\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eBrier_3yr\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ecoxph ★\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.885\u0026thinsp;\u0026plusmn;\u0026thinsp;0.117\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.022\u0026thinsp;\u0026plusmn;\u0026thinsp;0.038\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.043\u0026thinsp;\u0026plusmn;\u0026thinsp;0.051\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.197\u0026thinsp;\u0026plusmn;\u0026thinsp;0.382\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.525\u0026thinsp;\u0026plusmn;\u0026thinsp;0.093\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.535\u0026thinsp;\u0026plusmn;\u0026thinsp;0.022\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e \u003cp\u003e0.545\u0026thinsp;\u0026plusmn;\u0026thinsp;0.091\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003erfsrc ★\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.800\u0026thinsp;\u0026plusmn;\u0026thinsp;0.169\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.705\u0026thinsp;\u0026plusmn;\u0026thinsp;0.418\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.897\u0026thinsp;\u0026plusmn;\u0026thinsp;0.179\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.827\u0026thinsp;\u0026plusmn;\u0026thinsp;0.208\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.330\u0026thinsp;\u0026plusmn;\u0026thinsp;0.043\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.488\u0026thinsp;\u0026plusmn;\u0026thinsp;0.115\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e \u003cp\u003e0.548\u0026thinsp;\u0026plusmn;\u0026thinsp;0.085\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eridge\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.784\u0026thinsp;\u0026plusmn;\u0026thinsp;0.184\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.787\u0026thinsp;\u0026plusmn;\u0026thinsp;0.084\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.939\u0026thinsp;\u0026plusmn;\u0026thinsp;0.105\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.864\u0026thinsp;\u0026plusmn;\u0026thinsp;0.180\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.392\u0026thinsp;\u0026plusmn;\u0026thinsp;0.048\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.593\u0026thinsp;\u0026plusmn;\u0026thinsp;0.143\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e \u003cp\u003e0.678\u0026thinsp;\u0026plusmn;\u0026thinsp;0.112\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003exgboost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.768\u0026thinsp;\u0026plusmn;\u0026thinsp;0.228\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.773\u0026thinsp;\u0026plusmn;\u0026thinsp;0.321\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.835\u0026thinsp;\u0026plusmn;\u0026thinsp;0.164\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.794\u0026thinsp;\u0026plusmn;\u0026thinsp;0.179\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.251\u0026thinsp;\u0026plusmn;\u0026thinsp;0.025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.308\u0026thinsp;\u0026plusmn;\u0026thinsp;0.086\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e \u003cp\u003e0.334\u0026thinsp;\u0026plusmn;\u0026thinsp;0.083\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003estepwise\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.682\u0026thinsp;\u0026plusmn;\u0026thinsp;0.251\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNaN\u0026thinsp;\u0026plusmn;\u0026thinsp;NA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.833\u0026thinsp;\u0026plusmn;\u0026thinsp;NA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.900\u0026thinsp;\u0026plusmn;\u0026thinsp;0.141\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.770\u0026thinsp;\u0026plusmn;\u0026thinsp;0.315\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.840\u0026thinsp;\u0026plusmn;\u0026thinsp;0.219\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e \u003cp\u003e0.870\u0026thinsp;\u0026plusmn;\u0026thinsp;0.179\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eplsRcox\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.660\u0026thinsp;\u0026plusmn;\u0026thinsp;0.255\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.510\u0026thinsp;\u0026plusmn;\u0026thinsp;0.178\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.732\u0026thinsp;\u0026plusmn;\u0026thinsp;0.330\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.683\u0026thinsp;\u0026plusmn;\u0026thinsp;0.399\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.444\u0026thinsp;\u0026plusmn;\u0026thinsp;0.068\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.459\u0026thinsp;\u0026plusmn;\u0026thinsp;0.166\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e \u003cp\u003e0.431\u0026thinsp;\u0026plusmn;\u0026thinsp;0.156\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003elasso\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.575\u0026thinsp;\u0026plusmn;\u0026thinsp;0.107\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.795\u0026thinsp;\u0026plusmn;\u0026thinsp;0.289\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.807\u0026thinsp;\u0026plusmn;\u0026thinsp;0.269\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.608\u0026thinsp;\u0026plusmn;\u0026thinsp;0.193\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.306\u0026thinsp;\u0026plusmn;\u0026thinsp;0.077\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.350\u0026thinsp;\u0026plusmn;\u0026thinsp;0.159\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e \u003cp\u003e0.349\u0026thinsp;\u0026plusmn;\u0026thinsp;0.168\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eelastic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.545\u0026thinsp;\u0026plusmn;\u0026thinsp;0.077\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.545\u0026thinsp;\u0026plusmn;\u0026thinsp;0.064\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.559\u0026thinsp;\u0026plusmn;\u0026thinsp;0.261\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.574\u0026thinsp;\u0026plusmn;\u0026thinsp;0.220\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.279\u0026thinsp;\u0026plusmn;\u0026thinsp;0.064\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.349\u0026thinsp;\u0026plusmn;\u0026thinsp;0.173\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e \u003cp\u003e0.361\u0026thinsp;\u0026plusmn;\u0026thinsp;0.186\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003esurvivalsvm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.500\u0026thinsp;\u0026plusmn;\u0026thinsp;0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.168\u0026thinsp;\u0026plusmn;\u0026thinsp;0.020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.118\u0026thinsp;\u0026plusmn;\u0026thinsp;0.102\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.118\u0026thinsp;\u0026plusmn;\u0026thinsp;0.102\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.410\u0026thinsp;\u0026plusmn;\u0026thinsp;0.042\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.620\u0026thinsp;\u0026plusmn;\u0026thinsp;0.135\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e \u003cp\u003e0.710\u0026thinsp;\u0026plusmn;\u0026thinsp;0.096\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eWaterfall and force plots for individual samples demonstrated how each gene's expression pushed the model's prediction towards a high or low risk score (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eD). The waterfall plot showed the model predicts a high-risk score of f(x)\u0026thinsp;=\u0026thinsp;9.13 for this sample, significantly above the baseline expected value E\u003csup\u003e[f(x)]\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;2.63. The explanation revealed that 43 genetic features collectively contribute to this elevated risk. The most substantial positive contributions come from a subset of features, with EIF4EBP1 contributing\u0026thinsp;+\u0026thinsp;8.83 SHAP value units. These large positive values indicated that the expression pattern of the specific genes in this sample was the primary driver of its high-risk classification.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Development and validation of a robust prognostic model based on prognostic genes\u003c/h2\u003e \u003cp\u003eThe RSF model with the 48 candidate genes was used to calculate a risk score. Eight genes with an importance score\u0026thinsp;\u0026gt;\u0026thinsp;0.025 (EIF4EBP1, NRTN, COL9A3, TRIM63, FABP7, ALAD, H4C13, PLAU) were identified as the prognostic genes. Patients in both the TCGA training and GSE135565 validation cohorts were stratified into HRG and LRG using the optimal cut-points (TCGA-TNBC: 3.109546, GSE135565: 5.751818). K-M analysis confirmed that patients in the HRG had significantly worse OS than those in the LRG in both cohorts (log-rank p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA). The distribution of risk scores, survival status, and expression heatmap of the eight prognostic genes further validated the model's stratification power (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB). Time-dependent ROC analysis showed that the risk score had predictive value for 1-, 2-, and 3-year OS, with AUCs consistently above 0.8 in both cohorts (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Risk score as an independent prognostic factor\u003c/h2\u003e \u003cp\u003eUnivariate Cox regression identified the risk score, N stage, and M stage as potential prognostic factors (Cox p\u0026thinsp;\u0026lt;\u0026thinsp;0.2, HR\u0026thinsp;\u0026ne;\u0026thinsp;1, PH test p\u0026thinsp;\u0026gt;\u0026thinsp;0.05) (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA-B). Subsequent multivariate Cox regression confirmed that the risk score and M stage were independent prognostic factors for OS in TNBC (Cox p\u0026thinsp;\u0026lt;\u0026thinsp;0.2, HR\u0026thinsp;\u0026ne;\u0026thinsp;1, PH test p\u0026thinsp;\u0026gt;\u0026thinsp;0.05) (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eC-D). These results underscore the clinical relevance of the TMRG-based signature beyond conventional clinicopathological variables.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Immune infiltration and pathway activity in risk groups\u003c/h2\u003e \u003cp\u003eThe ssGSEA revealed significant differences in immune cell infiltration between risk groups (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eA-B). Central memory CD8\u003csup\u003e+\u003c/sup\u003e T cells, immature dendritic cells, and neutrophils were significantly enriched in the HRG (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Spearman correlation analysis showed that PLAU expression was significantly positively correlated with these immune cells (cor\u0026thinsp;\u0026gt;\u0026thinsp;0.3, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01), while NRTN was significantly negatively correlated with central memory CD8 T cells and neutrophils (cor \u0026lt; -0.4, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01) (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eC). The co-enrichment of these immune cells, coupled with their divergent correlations with PLAU and NRTN, suggests a complex immunomodulatory role for these prognostic genes, potentially contributing to an immunosuppressive microenvironment in high-risk TNBC.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eGSVA highlighted significant pathway activity differences between risk groups. Seven KEGG pathways were significantly enriched (|t| \u0026gt; 2, p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eD). Among these, five pathways were significantly activated in the HRG, including ''ABC transporters'', ''N glycan biosynthesis'', ''type II diabetes mellitus'', ''riboflavin metabolism'', and ''endocytosis''. Conversely, pathways including ''cysteine and methionine metabolism'' and ''glycosaminoglycan biosynthesis keratan sulfate'' were suppressed. The specific activation of ABC transporters and endocytosis in high-risk TNBC reveals a coordinated mechanism for drug efflux and membrane trafficking that likely contributes to chemoresistance, while the altered glycosylation and metabolic patterns suggest extensive reprogramming of cell surface properties and redox homeostasis that collectively promote tumor survival and progression.\u003c/p\u003e \u003cp\u003eTo elucidate the biological pathways distinguishing high- from low-risk TNBC patients, GSEA was performed using KEGG gene sets. A total of 19 pathways were significantly enriched (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05, |NES| \u0026gt; 1). The top five significantly enriched pathways in the HRG included ''steroid hormone biosynthesis'', ''pentose and glucuronate interconversions'', ''drug metabolism-other enzymes'', ''retinol metabolism'', and ''ascorbate and aldarate metabolism'' (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eE). These findings indicate that high-risk TNBC is characterized by pronounced metabolic reprogramming involving hormone metabolism, xenobiotic processing, and oxidative stress-related pathways, which may contribute to its aggressive phenotype and therapeutic resistance.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e3.6 Somatic mutational landscape and prognostic value of TMB\u003c/h2\u003e \u003cp\u003eThe waterfall diagram illustrated the somatic mutational profile across 91 TNBC samples, with 88 samples (96.7%) exhibiting at least one genetic alteration (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eA). TP53 demonstrated the highest mutation frequency, present in approximately 84% of samples, followed by TTN (19%) and MUC16 (14%). The mutation spectrum was dominated by missense mutations across all major genes.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eNo significant difference in TMB was observed between risk groups (p\u0026thinsp;=\u0026thinsp;0.858) (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eB). However, high TMB was associated with improved OS (p\u0026thinsp;=\u0026thinsp;0.032) (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eC). Patients with low risk and high TMB had the most favorable prognosis (p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001) (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eD), highlighting the complementary value of genomic and transcriptomic biomarkers.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section2\"\u003e \u003ch2\u003e3.7 AI-predicted therapeutic compounds\u003c/h2\u003e \u003cp\u003eGraphBAN analysis predicted high-affinity interactions between four prognostic genes (H4C13, COL9A3, ALAD, TRIM63) and two compounds: omega-3-carboxylic acids (ω-3 carboxylic acid) and sucralfate (interaction probability\u0026thinsp;\u0026gt;\u0026thinsp;0.8) (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003eA). Among the prognostic genes, COL9A3 was found to have the highest number of connecting edges, suggesting its higher potential as a novel drug target. Conversely, among the compounds, ω-3 carboxylic acid exhibited the most connections, indicating their strong potential as a novel therapeutic agent. No corresponding compounds were predicted for the remaining four prognostic genes. These findings suggest repurposing opportunities for existing drugs in TNBC treatment.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePrediction results from GraphBAN\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSMILES\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eProtein\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003epred\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDrug Name\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eGene\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCC/...O)O\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMSG...GGF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.800178885\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eOMEGA-3-CARBOXYLIC ACIDS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eH4C13\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCC/...O)O\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAG...RSS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.815427303\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eOMEGA-3-CARBOXYLIC ACIDS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eCOL9A3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCC/...O)O\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMQP...KEE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.889059484\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eOMEGA-3-CARBOXYLIC ACIDS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eALAD\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCC/...O)O\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMDY...GHQ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.823631704\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eOMEGA-3-CARBOXYLIC ACIDS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTRIM63\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eO.O...O)O\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAG...RSS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.806707144\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSUCRALFATE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eCOL9A3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003e3.8 Different drug sensitivity between HRG and LRG\u003c/h2\u003e \u003cp\u003eTo evaluate the potential clinical utility of the prognostic model in guiding treatment selection, the IC\u003csub\u003e50\u003c/sub\u003e values of 138 chemotherapeutic and targeted agents were estimated in the TCGA-TNBC cohort. Wilcoxon test identified ten drugs with significantly different IC\u003csub\u003e50\u003c/sub\u003e values between risk groups (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Among these, the majority exhibited lower IC\u003csub\u003e50\u003c/sub\u003e values in the LRG, including AZ628, indicating greater drug sensitivity (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003eB). These results suggest that patients in the LRG may be more responsive to conventional chemotherapeutic agents, highlighting the potential utility of the prognostic model in refining treatment strategies.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e3.9 Identification of key cells\u003c/h2\u003e \u003cp\u003eAfter quality control (\u003cb\u003eSupplementary Fig.\u0026nbsp;1A\u003c/b\u003e), the top 3,000 HVGs and 20 PCs were selected for cell clustering (\u003cb\u003eSupplementary Fig.\u0026nbsp;1B-D\u003c/b\u003e). Unsupervised clustering with resolution parameter 2 identified 53 distinct cell clusters (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003eA), which were annotated into 8 major cell types: B cells, common myeloid progenitors (CMPs), dendritic cells (DCs), endothelial cells, epithelial cells, macrophages, stromal cells, and T cells (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003eB). Differential abundance analysis revealed six cell types with significantly different abundances between TNBC and control tissues (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003eC). Stromal cells were significantly decreased in TNBC, while five immune cell types (T cells, B cells, macrophages, DCs, CMPs) were enriched, indicating immune activation in the TNBC microenvironment. UMAP visualization showed distinct expression patterns for prognostic genes (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003eD). Subsequently, differential expression analysis showed T cells, B cells, macrophages, and stromal cells exhibited significant differential expression of at least one prognostic gene and were therefore defined as the four key cell types for further analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003eE). DCs showed no significant differences, and CMPs were excluded as they were only present in TNBC samples.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003e3.10 Subcluster analysis and pseudotemporal trajectory of key cell types\u003c/h2\u003e \u003cp\u003eSubsequent secondary dimensionality reduction and clustering were performed on the four key cell types (macrophages, T cells, B cells, and stromal cells). Macrophages were subdivided into 14 distinct subclusters, T cells into 12 subclusters, B cells into 14 subclusters, and stromal cells into 9 subclusters (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003eA), revealing substantial heterogeneity within each major cellular compartment.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003ePseudotemporal trajectory analysis was further conducted to reconstruct the potential differentiation pathways or cellular state transitions within each of these key cell types. Each cell type exhibited a unique branched trajectory, visualized as a continuum from a presumed initial state (dark blue) towards a terminal state (light orange) along the inferred pseudotime (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003eB).\u003c/p\u003e \u003cp\u003eAnalysis of genes demonstrating significant dynamic expression changes during these transitions highlighted key mediators specific to each lineage. Notably, CCL20 in macrophages, CCL4 in T cells, FCER1G in B cells, and ACTG2 in stromal cells showed pronounced expression variations along their respective trajectories, as illustrated in expression heatmaps and trend plots (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003eC-D).\u003c/p\u003e \u003cp\u003eThe dynamic expression patterns of the prognostic genes were specifically examined along these pseudotemporal axes (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003eE). Distinct temporal expression programs were observed. In macrophages, FABP7 expression was biased towards the beginning of the pseudotime, whereas PLAU expression was enriched towards the end. For T cells, FABP7 exhibited higher expression during the early-to-mid phases of the differentiation trajectory. During B-cell differentiation, both FABP7 and PLAU showed elevated expression in the early phase. Stromal cells displayed a more complex pattern, with dynamic expression observed for EIF4EBP1, COL9A3, FABP7, and PLAU. Specifically, EIF4EBP1 and PLAU maintained relatively high expression throughout the early, middle, and late stages, COL9A3 was highly expressed in the early phase, and FABP7 expression increased during the later phase. These findings delineate cell-type-specific and differentiation-stage-dependent expression dynamics of the prognostic genes within the TNBC tumor microenvironment, suggesting their potential roles in distinct biological processes across different cellular contexts.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec27\" class=\"Section2\"\u003e \u003ch2\u003e3.11 TF activity and metabolic pathway activity\u003c/h2\u003e \u003cp\u003eThe TF activity analysis provided crucial insights into the regulatory programs operating in key cells of TNBC (Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003eA). The consistent activation of MYC and E2F4, coupled with suppression of ATF3 and HBP1, suggested coordinated pro-tumorigenic regulatory networks across different cell types.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eMetabolic scoring indicated elevated activity in pathways such as glycolysis/gluconeogenesis and fatty acid elongation in macrophages compared to other cell types (Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003eB). This aligns with the known metabolic plasticity of tumor-associated macrophages in TNBC.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003e3.12 Intercellular communication networks\u003c/h2\u003e \u003cp\u003eIn TNBC samples, B cells, T cells, and macrophages showed enhanced interaction intensity compared with controls, whereas stromal cells displayed stronger interactions in control tissues (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003eA-B). Analysis of ligand-receptor pair contributions revealed that macrophage-derived SPP1 signaling via the CD44 receptor complex was specifically activated in TNBC (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003eA-B). Collectively, these findings reveal a rewired intercellular communication landscape in TNBC, with active participation of immune and stromal cells in tumor-promoting signaling networks.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4 Discussion","content":"\u003cp\u003eTriple-negative breast cancer (TNBC) is an aggressive malignancy with limited therapies, poor prognosis, and a lack of reliable biomarkers. Tryptophan metabolism\u0026mdash;particularly the kynurenine pathway\u0026mdash;can promote an immunosuppressive tumor microenvironment (TME), though its prognostic role in TNBC remains unclear. Using multi-omics data, we identified 48 tryptophan metabolism-related genes by differential expression and univariate Cox analysis. An optimized random survival forest (RSF) model was built and validated as an 8-gene prognostic signature, including EIF4EBP1 and NRTN, which SHAP analysis highlighted as key risk stratification factors. High-risk TNBC showed an \"infiltrated but suppressed\" immune profile, with enriched central memory CD8⁺ T cells, immature dendritic cells, and neutrophils, plus activated oncogenic pathways like ABC transporters and endocytosis. The GraphBAN AI model predicted ω-3 carboxylic acid and sucralfate as potential therapeutics. scRNA-seq revealed cell-specific expression and differentiation trajectories of prognostic genes in T cells, B cells, macrophages, and stromal cells. This study offers a tryptophan metabolism-derived prognostic tool for TNBC, clarifies TME remodeling mechanisms, and suggests translational drug candidates to advance precision oncology.\u003c/p\u003e \u003cp\u003eA key finding is the development and validation of a robust prognostic model based on eight genes (EIF4EBP1, NRTN, COL9A3, TRIM63, FABP7, ALAD, H4C13, PLAU). The model showed strong predictive performance in both the TCGA training and GEO validation sets, and its risk score independently predicted overall survival and surpassed certain conventional TNM staging criteria, thereby highlighting its added clinical value, and these results imply that tryptophan metabolic reprogramming plays a more central role in TNBC progression than previously appreciated, reflecting tumor intrinsic aggressiveness(Yan et al. 2024, Xue et al. 2023a). Using machine learning, Random Survival Forest (RSF) was selected as the optimal model for its predictive accuracy and ability to capture complex gene\u0026ndash;gene interactions, consistent with the polygenic basis of tumorigenesis. SHAP analysis enhanced interpretability, identifying EIF4EBP1 and NRTN as the top contributors to high- and low-risk stratification, respectively, offering clear mechanistic targets; the high-risk gene EIF4EBP1, a downstream effector of mTORC1, regulates protein translation and cancer metabolism(Wu et al. \u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e2024\u003c/span\u003e, Fang et al. 2022). Under nutrient-rich conditions, mTORC1 phosphorylates and inhibits EIF4EBP1, promoting oncogenic protein synthesis. Consistent with prior reports(Wu et al. \u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), high EIF4EBP1 expression correlated with poor prognosis, possibly indicating sustained mTOR activation or non-canonical pro-tumor functions(Guo et al. 2020). In contrast, In contrast, NRTN, a GDNF family member mainly linked to neuronal survival(Correia et al. 2021), has an unclear role in TNBC; we propose that it may interact via GFRα2 with integrins on tumor or immune cells, potentially modulating ECM remodeling and pro-invasive signaling. Although it has an unclear role in TNBC, and we propose that NRTN it may interact via GFRα2 with integrins on tumor or immune cells to, potentially modulating ECM remodeling and pro-invasive signaling\u0026mdash;providing new insight into neurotrophic factors in cancer(Man et al. 2023). Integrated Bioinformatics analysis revealed significant enrichment of the eight genes are enriched in key TNBC pathways\u0026mdash;including (PI3K-Akt signaling, ECM\u0026ndash;receptor interaction, and focal adhesion). This suggesting a functional coupling between tryptophan metabolic dysregulation and core oncogenic networks, collectively influencing tumor malignancy and the immune microenvironment. Together, these findings support the biological relevance of our prognostic model.\u003c/p\u003e \u003cp\u003eIn this study, ssGSEA was employed to reveal a distinct immune infiltration landscape in high-risk TNBC, characterized by significant enrichment of central memory CD8⁺ T cells, immature dendritic cells, and neutrophils. Correlation analysis demonstrated that PLAU expression was significantly positively correlated with the infiltration of these immune cell subsets, whereas NRTN showed a significant negative correlation with central memory CD8⁺ T cells and neutrophils. This landscape presents a paradoxical scenario\u0026mdash;extensive immune cell infiltration without a corresponding survival benefit\u0026mdash;suggesting an 'infiltrated but suppressed' phenotype in high-risk TNBC. We hypothesize that PLAU-mediated extracellular matrix remodeling and chemotaxis may act as key drivers of immune cell recruitment, while functional suppression arises from the synergistic interplay of multiple mechanisms. First, tryptophan metabolism via the kynurenine pathway promotes T cell exhaustion and regulatory T cell differentiation(Man et al. 2023), which can lead to disrupted T cell receptor signaling and upregulation of co-inhibitory molecules, driving CD8⁺ T cells into a state of \"pseudo-exhaustion\" characterized by retained migratory capacity but loss of effector function(Koushki et al. 2021, Anderson et al. 2017, Scarlett et al. 2012). Second, the accumulation of immature dendritic cells, which exhibit low antigen-presenting capacity, may further reinforce immunosuppression by inducing T cell anergy(Scarlett et al. 2012). In addition, tumor-associated neutrophils can directly suppress T cell function through the secretion of arginase-1 and reactive oxygen species(Liu et al. 2022). Collectively, these three layers form a multidimensional immunosuppressive network that accounts for the paradoxical association between extensive immune infiltration and poor prognosis. The negative correlation between NRTN and immune cell infiltration further suggests that this protective gene may exert its effects, at least in part, by constraining such dysfunctional immune recruitment, although the underlying mechanisms warrant experimental investigation.\u003c/p\u003e \u003cp\u003eBeyond the immunosuppressive microenvironment features of high-risk TNBC highlighted earlier, molecular pathway dysregulation represents another critical driver of its aggressive phenotype and poor prognosis. GSVA of TNBC molecular heterogeneity revealed significant pathway activity differences between risk groups, offering insights into tumor biology and therapeutic opportunities. The high-risk group exhibited marked activation of pathways linked to substance transport, energy metabolism, and membrane function\u0026mdash;including \"ABC transporters,\" \"N-glycan biosynthesis,\" \"Type II diabetes,\" \"Riboflavin metabolism,\" and \"Endocytosis.\" These activated pathways collectively drive TNBC aggressiveness, therapy resistance, and poor prognosis. Specifically, ABC transporter upregulation in high-risk TNBC enhances drug efflux, reducing intracellular chemotherapeutic concentrations and driving multidrug resistance. Complementing the immunosuppressive features observed in high-risk TNBC, GSVA-based pathway enrichment analyses further delineate the molecular underpinnings of its aggressive phenotype. The coordinated activation of ABC transporters and endocytosis pathways in the high-risk group points to concurrently enhanced drug efflux and membrane trafficking(Chen et al. 2020, Egea et al. 2022). Consistently, this group showed significantly higher IC₅₀ values for most chemotherapeutic agents, suggesting a broadly refractory phenotype that may underlie chemoresistance in high-risk TNBC. Such a resistance barrier likely represents a critical clinical determinant of poor prognosis and may be potentially overcome by combining ABC transporter inhibitors with chemotherapy(Chien et al. 2025). Activation of N-glycan biosynthesis may synergize with PLAU-mediated extracellular matrix remodeling to facilitate tumor invasion and immune evasion. One potential mechanism involves altered glycosylation of adhesion molecules such as integrins, which can modulate cell\u0026ndash;matrix interactions and immune recognition(Shi et al. \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2020\u003c/span\u003e, Xue et al. \u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e2023b\u003c/span\u003e). Moreover, glycosylated proteins derived from this pathway hold promise as diagnostic biomarkers or therapeutic targets(Ščup\u0026aacute;kov\u0026aacute; et al. \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Enrichment of the type 2 diabetes mellitus pathway aligns with the lipid metabolic reprogramming indicated by early high expression of FABP7 in the high-risk group, suggesting that concurrent dysregulation of glucose and lipid metabolism supports tumor proliferation and survival. Notably, exosome-mediated signaling under insulin-resistant conditions has been reported to promote epithelial\u0026ndash;mesenchymal transition(Qiu et al. \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), further linking metabolic disturbances to aggressive behavior. Conversely, suppression of cysteine and methionine metabolism, along with reduced keratan sulfate biosynthesis, suggests impaired glutathione synthesis and altered extracellular matrix sulfation. These changes may contribute to disease progression by increasing oxidative stress and promoting matrix remodeling(Park et al. 2022, Upadhyayula et al. 2023, Leiphrakpam et al. 2019), potentially creating a pro-tumorigenic microenvironment. Collectively, high-risk TNBC exhibits a synergistic pattern characterized by enhanced transport, metabolic reprogramming, and matrix dysregulation, which closely aligns with the multidimensional malignant features captured by the prognostic genes identified in this study. From a translational perspective, these findings highlight several candidate strategies for improving outcomes in high-risk patients, including combining ABC transporter inhibitors with chemotherapy to counter chemoresistance, targeting N-glycan biosynthesis pathways to potentially restore antitumor immunity, and modulating sulfur amino acid metabolism to disrupt metabolic adaptation.\u003c/p\u003e \u003cp\u003eOur single-cell transcriptomic analysis provided unprecedented resolution into the TNBC tumor microenvironment (TME). Beyond confirming its immune-rich, stroma-poor nature, we delineated cell type-specific expression of eight core prognostic genes in T cells, B cells, macrophages, and stromal cells. Pseudotime analysis further revealed the dynamic expression of these eight prognostic genes during key cellular state transitions. Notably, FABP7 and PLAU displayed distinct temporal expression patterns during macrophage differentiation: FABP7 peaked early, while PLAU was markedly upregulated later, thereby suggesting stage-specific functional specialization. As a key lipid metabolism regulator, FABP7 facilitates fatty acid uptake and transport, thereby supporting tumor survival and proliferation. FABP7 may also influence energy homeostasis and the cell cycle via the PPAR-α signaling pathway(Kwong et al. 2020, Kwong et al. 2019, Miyazaki et al. 2025, Xu et al. 2020). Its early expression likely supplies lipids and energy to drive M2 polarization, thereby fostering an immunosuppressive microenvironment(Miyazaki et al. 2025, Xu et al. 2025). In contrast, PLAU\u0026mdash;a central element of the plasminogen activation system\u0026mdash;promotes tumor invasion, angiogenesis, and the ECM degradation(Shi et al. \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Late PLAU upregulation may indicate macrophage maturation, thereby enabling local invasion and metastasis via ECM breakdown(Zhu et al. \u003cspan citationid=\"CR75\" class=\"CitationRef\"\u003e2021\u003c/span\u003e, Wan et al. 2025). These temporal patterns outline a functional progression: early metabolic remodeling via FABP7 supports polarization, while late PLAU expression facilitates pro-invasive ECM degradation. Therapeutically, early FABP7 inhibition may disrupt metabolic adaptation and M2 macrophage polarization, while late PLAU targeting may suppress TNBC metastasis. Additionally, the sustained high expression of EIF4EBP1 and PLAU in stromal cells indicates that stromal cells are highly activated cancer-associated fibroblasts (CAFs) with enhanced protein synthesis and ECM secretion\u0026mdash;consistent with reported fibroblast activation programs(Torrence et al. 2021) \u0026mdash;thereby corroborating, at single-cell resolution, the active stromal remodeling in high-risk TNBC.\u003c/p\u003e \u003cp\u003eBased on the cell-specific expression patterns of prognostic genes revealed by single-cell analysis, we further employed the GraphBAN model to screen for potential therapeutic compounds targeting these genes. A key translational finding of this study is the utilization of the GraphBAN model\u0026mdash;a graph neural network (GNN)\u0026mdash;for translating molecular discoveries into clinical therapy. This AI model systematically predicted interactions between ω-3 carboxylic acid and key prognostic components such as COL9A3. While a long-chain polyunsaturated fatty acid, ω-3 carboxylic acid, is known for its anti-inflammatory, antioxidant, and potential anti-tumor properties(Nicholls et al. 2020, Cikoš et al. 2021), its specific targets and mechanisms in TNBC remain unclear. Our model proposes a novel hypothesis: ω-3 carboxylic acid may exert anti-TNBC effects by binding specifically to COL9A3, a notion requiring experimental validation. COL9A3, a component of type IX collagen, is overexpressed in aggressive tumors and may promote malignancy by activating integrin-mediated FAK/Src and PI3K/Akt signaling, thereby inducing EMT and enhancing migration and invasion(Yang et al. \u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e2024\u003c/span\u003e, Wu et al. 2021b). Based on this, we speculate that ω-3 carboxylic acids may suppress the pro-tumor function of COL9A3 by interfering with its interaction with integrins, a mechanism that requires experimental validation. This insight suggests novel therapeutic avenues for TNBC via targeted ECM modulation. Notably, ω-3 carboxylic acid, an FDA-approved agent for hypertriglyceridemia, is increasingly studied in cancer for its anti-inflammatory and potential antitumor effects. Given the tryptophan metabolism-driven immunosuppressive microenvironment identified in our study, its immunomodulatory properties may further counterbalance this effect, offering a rationale for combination therapy. We therefore propose ω-3 carboxylic acid as a priority candidate for TNBC therapeutic development. Preclinical evaluation should confirm binding to COL9A3, assess monotherapy and combination therapy efficacy in animal models, and elucidate effects on the TME\u0026mdash;notably immune cell infiltration and function. This AI-guided approach underscores the potential of computational prediction to accelerate translational oncology.\u003c/p\u003e \u003cp\u003eThis study established and validated a robust tryptophan metabolism-related prognostic model for TNBC via integrated multi-omics analysis, systematically elucidating its biological basis at the pathway, TME, and functional levels: it correlates with PI3K-Akt signaling activation and ECM remodeling, reflects an \"infiltrated but suppressed\" immune state (characterized by tryptophan metabolic dysregulation, T cell pseudo-exhaustion, and neutrophil-mediated immunosuppression), and captures enhanced ABC transporter-mediated drug efflux plus metabolic adaptations of riboflavin metabolism and type II diabetes-related pathways. Additionally, single-cell transcriptomics and pseudotime analysis mapped prognostic genes to specific cellular subsets and revealed dynamic expression of key genes (e.g., FABP7, PLAU) during macrophage differentiation (yielding temporal-spatial functional insights), while the GraphBAN AI model identified potential interactions between approved drugs (e.g., ω-3 carboxylic acid) and core targets (e.g., COL9A3) to propose novel therapeutic candidates. A key limitation is its reliance on public datasets, which requires future experimental validation. Future work will: (1) functionally validate core genes (e.g., EIF4EBP1, NRTN) in TNBC malignancy; (2) elucidate their mechanisms in tryptophan metabolism-mediated immunosuppression; and (3) preclinically evaluate predicted drugs, including ω-3 carboxylic acid, alone and combined with chemotherapy or immunotherapy. In summary, this study provides a practical prognostic tool for TNBC, deepens multi-omics understanding of tryptophan metabolic reprogramming in progression, and offers translatable leads for targeted and combination immunotherapy.\u003c/p\u003e"},{"header":"5 Conclusion","content":"\u003cp\u003eOur integrated multi-omics study constructed and validated an eight-gene prognostic signature\u0026mdash;centered on tryptophan metabolism (EIF4EBP1, NRTN, COL9A3, etc.)\u0026mdash;for TNBC. The RSF model was identified as the optimal model, generating a risk score that independently stratified patient survival in both TCGA and GEO cohorts while outperforming conventional clinical predictors. Biologically, the signature reflects coordinated PI3K-Akt pathway activation and extracellular matrix (ECM) remodeling, and correlates with an immunosuppressive tumor microenvironment characterized by dysregulated tryptophan metabolism, T cell dysfunction, and neutrophil-mediated suppression. Single-cell analyses revealed differentiation-associated expression dynamics of FABP7 and PLAU in macrophages. GraphBAN AI modeling further predicted \u0026omega;-3 carboxylic acid interaction with COL9A3, proposing a new therapeutic avenue. Although experimental confirmation is needed, this work delivers a robust prognostic tool, enhances multi-omics understanding of tryptophan rewiring in TNBC, and suggests actionable targets for combination therapy.\u003c/p\u003e"},{"header":"Abbreviation","content":"\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003escRNA-seq\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSingle-Cell RNA Sequencing / Transcriptomics and RNA Sequencing\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eGEO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eGene Expression Omnibus\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eQC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eQuality Control\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePCs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrincipal Components\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUMAP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUniform Manifold Approximation and Projection\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"Declarations","content":"\u003cp\u003eData Availability Statement\u003c/p\u003e\n\u003cp\u003eThe datasets analysed in this study are available in The Cancer Genome Atlas (TCGA) repository (https://portal.gdc.cancer.gov/)，in Gene Expression Omnibus (GEO) database (https://www.ncbi. nlm.nih.gov/gds), including GSE135565 and GSE161529\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflict of\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003ei\u003c/strong\u003e\u003cstrong\u003enterest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eConceptualization: Youjun Wu; data curation: Xiaorong Pang; formal analysis: Feng Cen; investigation: Liang Xie, methodology: Youjun Wu; resources: Xianglan Mo; software: Xiang Feng; supervision: Xiaorong Pang; validation: Feng Cen; visualization: Xiang Feng; writing\u0026mdash;original draft: Youjun Wu; writing\u0026mdash;review and editing: Xianglan Mo. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was supported by the Guangxi Natural Science Foundation Project (grant number：2023GXNSFBA026013) and the Science and Technology Base and Talent Program in Guangxi (grant number：GuikeAD23026097).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe would like to express our sincere gratitude to all individuals and organizations who supported and assisted us throughout this research. Without your support, this study would not have been feasible.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for Publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;to patients\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eANDERSON K. G. et al. (2017) Obstacles Posed by the Tumor Microenvironment to T cell Activity: A Case for Synergistic Therapies. Cancer Cell 31(3):311-325\u003c/li\u003e\n \u003cli\u003eARAN D. et al. (2019) Reference-based analysis of lung single-cell sequencing reveals a transitional profibrotic macrophage. Nat Immunol 20(2):163-172\u003c/li\u003e\n \u003cli\u003eBRAY F. et al. (2024) Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 74(3):229-263\u003c/li\u003e\n \u003cli\u003eCAO J. et al. (2019) The single-cell transcriptional landscape of mammalian organogenesis. Nature 566(7745):496-502\u003c/li\u003e\n \u003cli\u003eCHANG C. H. et al. (2019) A novel orally available seleno-purine molecule suppresses triple-negative breast cancer cell proliferation and progression to metastasis by inducing cytostatic autophagy. Autophagy 15(8):1376-1390\u003c/li\u003e\n \u003cli\u003eCHAROENTONG P. et al. (2017) Pan-cancer Immunogenomic Analyses Reveal Genotype-Immunophenotype Relationships and Predictors of Response to Checkpoint Blockade. Cell Rep 18(1):248-262\u003c/li\u003e\n \u003cli\u003eCHEN H. \u0026amp; BOUTROS, P. C. (2011) VennDiagram: a package for the generation of highly-customizable Venn and Euler diagrams in R. BMC Bioinformatics 12:35\u003c/li\u003e\n \u003cli\u003eCHEN P. S. et al. (2020) Pathophysiological implications of hypoxia in human diseases. J Biomed Sci 27(1):63\u003c/li\u003e\n \u003cli\u003eCHENG Q. et al. (2021) Three hematologic/immune system-specific expressed genes are considered as the potential biomarkers for the diagnosis of early rheumatoid arthritis through bioinformatics analysis. J Transl Med 19(1):18\u003c/li\u003e\n \u003cli\u003eCHIEN P. Y. et al. (2025) 2\u0026apos;,6\u0026apos;-dihydroxy-3\u0026apos;,4\u0026apos;-dimethoxydihydrochalcone counteracts cancer multidrug resistance by impeding STAT3 activation and ABC transporter-mediated drug efflux. Biomed Pharmacother 188:118153\u003c/li\u003e\n \u003cli\u003eCIKO\u0026Scaron; A. M. et al. (2021) Bioprospecting of Coralline Red Alga Amphiroa rigida J.V. Lamouroux: Volatiles, Fatty Acids and Pigments. Molecules 26(3):520\u003c/li\u003e\n \u003cli\u003eCORREIA J. C. et al. (2021) Muscle-secreted neurturin couples myofiber oxidative metabolism and slow motor neuron identity. Cell Metab 33(11):2215-2230.e8\u003c/li\u003e\n \u003cli\u003eDOWLING C. M. et al. (2021) Multiple screening approaches reveal HDAC6 as a novel regulator of glycolytic metabolism in triple-negative breast cancer. Sci Adv 7(3): eabc4897\u003c/li\u003e\n \u003cli\u003eEGEA J. et al. (2022) Alkylating Agent-Induced Toxicity and Melatonin-Based Therapies. Front Pharmacol 13:873197\u003c/li\u003e\n \u003cli\u003eFANG Z. et al. (2022) Overactivation of hepatic mechanistic target of rapamycin kinase complex 1 (mTORC1) is associated with low transcriptional activity of transcription factor EB and lysosomal dysfunction in dairy cows with clinical ketosis. J Dairy Sci 105(5):4520-4533\u003c/li\u003e\n \u003cli\u003eGREENE L. I. et al. (2019) A Role for Tryptophan-2,3-dioxygenase in CD8 T-cell Suppression and Evidence of Tryptophan Catabolism in Breast Cancer Patient Plasma. Mol Cancer Res 17(1):131-139\u003c/li\u003e\n \u003cli\u003eGU Z. \u0026amp; H\u0026uuml;BSCHMANN, D. (2022) Make Interactive Complex Heatmaps in R. Bioinformatics 38(5):1460-1462\u003c/li\u003e\n \u003cli\u003eGUO B. et al. (2020) Micropeptide CIP2A-BP encoded by LINC00665 inhibits triple-negative breast cancer progression. Embo j 39(1):e102190\u003c/li\u003e\n \u003cli\u003eGUSTAVSSON E. K. et al. (2022) ggtranscript: an R package for the visualization and interpretation of transcript isoforms using ggplot2. Bioinformatics 38(15):3844-3846\u003c/li\u003e\n \u003cli\u003eHADIPOUR H. et al. (2025) GraphBAN: An inductive graph-based approach for enhanced prediction of compound-protein interactions. Nat Commun 16(1):2541\u003c/li\u003e\n \u003cli\u003eHANAHAN D. (2022) Hallmarks of Cancer: New Dimensions. Cancer Discov 12(1):31-46\u003c/li\u003e\n \u003cli\u003eH\u0026auml;NZELMANN S. et al. (2013) GSVA: gene set variation analysis for microarray and RNA-seq data. BMC Bioinformatics 14:7\u003c/li\u003e\n \u003cli\u003eHEAGERTY P. J. et al. (2000) Time-dependent ROC curves for censored survival data and a diagnostic marker. Biometrics 56(2):337-344\u003c/li\u003e\n \u003cli\u003eJAHCHAN N. S. et al. (2019) Tuning the Tumor Myeloid Microenvironment to Fight Cancer. Front Immunol 10:1611\u003c/li\u003e\n \u003cli\u003eJI P. et al. (2022) In vivo multidimensional CRISPR screens identify Lgals2 as an immunotherapy target in triple-negative breast cancer. Sci Adv 8(26):eabl8247\u003c/li\u003e\n \u003cli\u003eJIN S. et al. (2024) A prognostic model for overall survival in recurrent glioma patients treated with bevacizumab-containing therapy. Discov Oncol 15(1):85\u003c/li\u003e\n \u003cli\u003eKOUSHKI K. et al. (2021) Role of myeloid-derived suppressor cells in viral respiratory infections; Hints for discovering therapeutic targets for COVID-19. Biomed Pharmacother 144:112346\u003c/li\u003e\n \u003cli\u003eKUO L. W. et al. (2024) Blocking Tryptophan Catabolism Reduces Triple-Negative Breast Cancer Invasive Capacity. Cancer Res Commun 4(10):2699-2713\u003c/li\u003e\n \u003cli\u003eKWONG S. C. et al. (2020) Fatty acid binding protein 7 mediates linoleic acid-induced cell death in triple negative breast cancer cells by modulating 13-HODE. Biochimie 179:23-31\u003c/li\u003e\n \u003cli\u003eKWONG S. C. et al. (2019) Metabolic role of fatty acid binding protein 7 in mediating triple-negative breast cancer cell death via PPAR-\u0026alpha; signaling. J Lipid Res 60(11):1807-1817\u003c/li\u003e\n \u003cli\u003eLEI J. et al. (2023) Clinicopathological characteristics of pheochromocytoma/paraganglioma and screening of prognostic markers. J Surg Oncol 128(4):510-518\u003c/li\u003e\n \u003cli\u003eLEIPHRAKPAM P. D. et al. (2019) Role of keratan sulfate expression in human pancreatic cancer malignancy. Sci Rep 9(1):9665\u003c/li\u003e\n \u003cli\u003eLERCHER A. et al. (2020) Systemic Immunometabolism: Challenges and Opportunities. Immunity 53(3):496-509\u003c/li\u003e\n \u003cli\u003eLIN M. et al. (2025) Integrating multi-omics data of Triple-Negative Breast Cancer to explore the role of Kynurenine pathway and KYNU as a therapeutic target. Biochem Biophys Res Commun 756:151569\u003c/li\u003e\n \u003cli\u003eLIU P. et al. (2022) Pretreatment Systemic Immune-Inflammation Index Can Predict Response to Neoadjuvant Chemotherapy in Cervical Cancer at Stages IB2-IIB. Pathol Oncol Res 28:1610294\u003c/li\u003e\n \u003cli\u003eLIU P. et al. (2020) Potential Molecular Mechanisms of Plantain in the Treatment of Gout and Hyperuricemia Based on Network Pharmacology. Evid Based Complement Alternat Med 2020:30231273023127\u003c/li\u003e\n \u003cli\u003eLIU T. T. et al. (2021) Identification of CDK2-Related Immune Forecast Model and ceRNA in Lung Adenocarcinoma, a Pan-Cancer Analysis. Front Cell Dev Biol 9:682002\u003c/li\u003e\n \u003cli\u003eLOVE M. I. et al. (2014) Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol 15(12):550\u003c/li\u003e\n \u003cli\u003eMA P. et al. (2023) Bacterial droplet-based single-cell RNA-seq reveals antibiotic-associated heterogeneous cellular states. Cell 186(4):877-891.e14\u003c/li\u003e\n \u003cli\u003eMAN J. et al. (2023) TANGO1 interacts with NRTN to promote hepatocellular carcinoma progression by regulating the PI3K/AKT/mTOR signaling pathway. Biochem Pharmacol 213:115615\u003c/li\u003e\n \u003cli\u003eMIYAZAKI H. et al. (2025) FABP7 in Hepatic Macrophages Promotes Fibroblast Activation and CD4(+) T-Cell Migration by Regulating M2 Polarization During Liver Fibrosis. J Immunol Res 2025:6987981\u003c/li\u003e\n \u003cli\u003eNICHOLLS S. J. et al. (2020) Effect of High-Dose Omega-3 Fatty Acids vs Corn Oil on Major Adverse Cardiovascular Events in Patients at High Cardiovascular Risk: The STRENGTH Randomized Clinical Trial. Jama 324(22):2268-2280\u003c/li\u003e\n \u003cli\u003ePARK J. Y. et al. (2022) Dysbiotic change in gastric microbiome and its functional implication in gastric carcinogenesis. Sci Rep 12(1):4285\u003c/li\u003e\n \u003cli\u003eQIU Y. et al. (2025) Insulin Resistance Increases TNBC Aggressiveness and Brain Metastasis via Adipocyte-Derived Exosomes. Mol Cancer Res 23(6):567-578\u003c/li\u003e\n \u003cli\u003eRITCHIE M. E. et al. (2015) limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res 43(7):e47\u003c/li\u003e\n \u003cli\u003eROGERS T. J. et al. (2019) Reversal of Triple-Negative Breast Cancer EMT by miR-200c Decreases Tryptophan Catabolism and a Program of Immunosuppression. Mol Cancer Res 17(1):30-41\u003c/li\u003e\n \u003cli\u003eSATIJA R. et al. (2015) Spatial reconstruction of single-cell gene expression data. Nat Biotechnol 33(5):495-502\u003c/li\u003e\n \u003cli\u003eSCARLETT U. K. et al. (2012) Ovarian cancer progression is controlled by phenotypic changes in dendritic cells. J Exp Med 209(3):495-506\u003c/li\u003e\n \u003cli\u003e\u0026Scaron;ČUP\u0026aacute;KOV\u0026aacute; K. et al. (2021) Clinical importance of high-mannose, fucosylated, and complex N-glycans in breast cancer metastasis. JCI Insight 6(24): e146945\u003c/li\u003e\n \u003cli\u003eSHI K. et al. (2024) Pan-cancer analysis of PLAU indicates its potential prognostic value and correlation with neutrophil infiltration in BLCA. Biochim Biophys Acta Mol Basis Dis 1870(2):166965\u003c/li\u003e\n \u003cli\u003eSHI S. et al. (2020) Inhibition of MAN2A1 Enhances the Immune Response to Anti-PD-L1 in Human Tumors. Clin Cancer Res 26(22):5990-6002\u003c/li\u003e\n \u003cli\u003eSUGIAWAN Y. et al. (2023) Assessing the United Nations sustainable development goals from the inclusive wealth perspective. Sci Rep 13(1):1601\u003c/li\u003e\n \u003cli\u003eTAN Z. et al. (2025) Machine Learning and Experimental Validation Reveal MYH11 as a Novel Prognostic Biomarker and Therapeutic Target in Bladder Cancer. J Inflamm Res 18:8357-8387\u003c/li\u003e\n \u003cli\u003eTAN Z. et al. (2023) Comprehensive analysis of scRNA-Seq and bulk RNA-Seq reveals dynamic changes in the tumor immune microenvironment of bladder cancer and establishes a prognostic model. J Transl Med 21(1):223\u003c/li\u003e\n \u003cli\u003eTORRENCE M. E. et al. (2021) The mTORC1-mediated activation of ATF4 promotes protein and glutathione synthesis downstream of growth signals. Elife 10:e63326\u003c/li\u003e\n \u003cli\u003eUPADHYAYULA P. S. et al. (2023) Dietary restriction of cysteine and methionine sensitizes gliomas to ferroptosis and induces alterations in energetic metabolism. Nat Commun 14(1):1187\u003c/li\u003e\n \u003cli\u003eUTHAYOPAS K. et al. (2021) TSMDA: Target and symptom-based computational model for miRNA-disease-association prediction. Mol Ther Nucleic Acids 26:536-546\u003c/li\u003e\n \u003cli\u003eVEGHINI L. et al. (2024) Differential activity of MAPK signalling defines fibroblast subtypes in pancreatic cancer. Nat Commun 15(1):10534\u003c/li\u003e\n \u003cli\u003eWAN L. et al. (2025) Hypoxia-induced tumor cell-intrinsic PLAU activation drives immunotherapy resistance in collagenic lung adenocarcinoma. Int Immunopharmacol 162:115161\u003c/li\u003e\n \u003cli\u003eWANG Y. et al. (2020) LncRNA-encoded polypeptide ASRPS inhibits triple-negative breast cancer angiogenesis. J Exp Med http://doi.org/10.1084/jem.20190950\u003c/li\u003e\n \u003cli\u003eWU T. et al. (2021a) clusterProfiler 4.0: A universal enrichment tool for interpreting omics data. Innovation (Camb) 2(3):100141\u003c/li\u003e\n \u003cli\u003eWU W. et al. (2021b) Drivers and suppressors of triple-negative breast cancer. Proc Natl Acad Sci U S A 118(33): e2104162118\u003c/li\u003e\n \u003cli\u003eWU X. et al. (2024) Tumoral EIF4EBP1 regulates the crosstalk between tumor-associated macrophages and tumor cells in MRTK. Eur J Pharmacol 978:176787\u003c/li\u003e\n \u003cli\u003eXU D. et al. (2020) Upregulation of FABP7 inhibits acute kidney injury-induced TCMK-1 cell apoptosis via activating the PPAR gamma signalling pathway. Mol Omics 16(6):533-542\u003c/li\u003e\n \u003cli\u003eXU S. et al. (2025) FABP7-mediated lipid-laden macrophages drive the formation of pre-metastatic niche and liver metastasis. Int J Biol Sci 21(10):4388-4409\u003c/li\u003e\n \u003cli\u003eXUE C. et al. (2023a) Tryptophan metabolism in health and disease. Cell Metab 35(8):1304-1326\u003c/li\u003e\n \u003cli\u003eXUE Z. et al. (2023b) Investigation on acquired palbociclib resistance by LC-MS based multi-omics analysis. Front Mol Biosci 10:1116398\u003c/li\u003e\n \u003cli\u003eYAN J. et al. (2024) Molecular mechanisms and therapeutic significance of Tryptophan Metabolism and signaling in cancer. Mol Cancer 23(1):241\u003c/li\u003e\n \u003cli\u003eYANG X. et al. (2024) Potential regulation and prognostic model of colorectal cancer with extracellular matrix genes. Heliyon 10(16):e36164\u003c/li\u003e\n \u003cli\u003eYU L. et al. (2022) A Retrospective and Multicenter Study on COVID-19 in Inner Mongolia: Evaluating the Influence of Sampling Locations on Nucleic Acid Test and the Dynamics of Clinical and Prognostic Indexes. Front Med (Lausanne) 9:830484\u003c/li\u003e\n \u003cli\u003eZHANG D. et al. (2022a) Novel insight on marker genes and pathogenic peripheral neutrophil subtypes in acute pancreatitis. Front Immunol 13:964622\u003c/li\u003e\n \u003cli\u003eZHANG D. et al. (2022b) Facile synthesis of near-infrared responsive on-demand oxygen releasing nanoplatform for precise MRI-guided theranostics of hypoxia-induced tumor chemoresistance and metastasis in triple negative breast cancer. J Nanobiotechnology 20(1):104\u003c/li\u003e\n \u003cli\u003eZHANG Z. et al. (2023) Mitochondrial energy metabolism correlates with an immunosuppressive tumor microenvironment and poor prognosis in esophageal squamous cell carcinoma. Comput Struct Biotechnol J 21:4118-4133\u003c/li\u003e\n \u003cli\u003eZHOU F. et al. (2021) Photo-activated chemo-immunotherapy for metastatic cancer using a synergistic graphene nanosystem. Biomaterials 265:120421\u003c/li\u003e\n \u003cli\u003eZHU T. et al. (2021) Identification of a Competing Endogenous RNA Network Related to Immune Signature in Lung Adenocarcinoma. Front Genet 12:665555\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Triple-negative breast cancer, Tryptophan metabolism, Machine learning, Prognostic model, Single-cell RNA sequencing","lastPublishedDoi":"10.21203/rs.3.rs-9287896/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9287896/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground: \u003c/strong\u003eTriple-negative breast cancer (TNBC) is an aggressive malignancy with limited therapeutic options and poor prognosis. The kynurenine pathway of tryptophan metabolism contributes to an immunosuppressive tumor microenvironment (TME); however, its prognostic significance and molecular mechanisms in TNBC require further multi-omics characterization.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003eBulk and single-cell RNA-seq datasets were obtained from public repositories. Differentially expressed tryptophan metabolism-related genes were identified and screened via univariate Cox regression. Nine machine learning algorithms were trained using 5-fold cross-validation, with SHAP analysis applied for model interpretability. A prognostic risk model was developed, externally validated, and further analyzed for immune infiltration, pathway activity, and AI-driven drug screening. ScRNA-seq data were used to identify key cell populations and differentiation trajectories.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eAmong 48 candidates, Random Survival Forest demonstrated optimal performance and was selected to construct an eight-gene prognostic signature (EIF4EBP1, NRTN, COL9A3, TRIM63, FABP7, ALAD, H4C13, PLAU). The model effectively stratified patients into high- and low-risk groups with significantly distinct survival outcomes across multiple cohorts. High-risk patients exhibited increased infiltration of central memory CD8+ T cells, immature dendritic cells, and neutrophils, along with upregulated ABC transporter and endocytosis pathways. AI-driven screening identified omega-3-carboxylic acids and sucralfate as potential therapeutics. ScRNA-seq revealed that prognostic markers were predominantly expressed in T cells, B cells, macrophages, and stromal cells, with dynamic changes along differentiation trajectories.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion: \u003c/strong\u003eThis study establishes a novel tryptophan metabolism-derived prognostic signature for TNBC, providing insights into TME remodeling and identifying potential therapeutic strategies for high-risk patients.\u003c/p\u003e","manuscriptTitle":"Integrated bulk and single-cell transcriptomic analysis reveals a tryptophan metabolism-driven prognostic signature and therapeutic landscape in triple- negative breast cancer","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-03 11:20:14","doi":"10.21203/rs.3.rs-9287896/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"d5ec851f-ebe0-4f5a-83cf-ea9b2f70c48b","owner":[],"postedDate":"April 3rd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-14T16:25:22+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-03 11:20:14","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9287896","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9287896","identity":"rs-9287896","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}