Development of T-Cell-Associated Signatures Based on Machine Learning Integration for Predicting Breast Cancer Prognosis

Technical Summary: Development of T-Cell-Associated Signatures Based on Machine Learning Integration for Predicting Breast Cancer Prognosis

Problem Statement
Breast cancer remains the most prevalent cancer globally, with prognosis significantly deteriorating upon metastasis, particularly to bone. While single-cell sequencing has elucidated tumor heterogeneity and the tumor microenvironment (TME), the specific role of T-cell dynamics in breast cancer progression and prognosis remains incompletely understood. Furthermore, existing clinical staging and molecular typing often fail to fully capture the complexity of tumor-immune interactions, necessitating more robust biomarkers for prognostic prediction and immunotherapy response assessment.

Methodology
The study employed a multi-omics integration strategy combining single-cell RNA sequencing (scRNA-seq) and bulk RNA-seq data, analyzed through advanced machine learning pipelines:

1. Data Acquisition and Preprocessing:

   • scRNA-seq: Two datasets (GSE180286 and GSE158399) containing primary breast cancer tissues and paired lymph node metastases were analyzed using the R package Seurat. Batch effects were removed via SCTransform, and dimensionality reduction (PCA, t-SNE) was performed.
   • Bulk RNA-seq: Transcriptome and clinical data were retrieved from The Cancer Genome Atlas (TCGA-BRCA) and six independent GEO datasets (GSE1456, GSE20685, GSE24450, GSE42568, GSE58812, GSE7390).
   • Immunotherapy Cohorts: Validation was conducted using three immunotherapy datasets: GSE78220 (melanoma), GSE135222 (small cell lung cancer), and IMvigor210 (urothelial cancer).

2. Signature Construction:

   • Cell Annotation & Differential Analysis: scRNA-seq data identified 13 cell clusters. T-cell populations showed significant differential abundance between tumor and metastatic tissues. Differential genes (logFC > 0.5, p < 0.05) between these T-cell states were extracted.
   • Subtyping: Consensus clustering based on T-cell differential genes divided TCGA-BRCA samples into two subtypes (C1 and C2).
   • Machine Learning Integration: From the TCGA cohort, 419 differential genes were screened via univariate Cox regression to identify 52 prognostic genes. These were subjected to 101 combinations of 10 machine learning algorithms (including Random Survival Forest, Elastic Net, Lasso, Ridge, Stepwise Cox, CoxBoost, etc.).
   • Optimal Model Selection: The model with the highest average Harrell's concordance index (C-index) across datasets was selected. The optimal combination was identified as StepCox (forward) + Ridge regression.

3. Validation and Analysis:

   • Prognostic Value: Patients were stratified into high- and low-risk groups based on the median risk score. Survival analysis (Kaplan-Meier), ROC curves, and C-index comparisons were performed across all datasets.
   • TME and Multi-omics: Immune infiltration was assessed using ssGSEA, ESTIMATE, and CIBERSORT. Somatic mutations were analyzed via maftools to calculate Tumor Mutation Burden (TMB). The Tumor Immune Dysfunction and Exclusion (TIDE) score was computed to predict immunotherapy response.

Key Results

• Identification of T-Cell Subtypes: Two consensus clusters (C1 and C2) were identified. C2 exhibited a "hot" tumor phenotype with significantly higher immune cell infiltration (including activated CD8+ T cells, NK cells, and macrophages) and a significantly better overall survival rate compared to C1.
• Robustness of the TRS Model: The T-cell-associated signature (TRS), derived from the StepCox+Ridge model, demonstrated superior prognostic accuracy compared to conventional clinical traits (age, sex, T/N/M staging, and molecular subtype) across seven independent datasets. The model maintained high accuracy with C-index values ranging from 0.682 to 0.769 in various cohorts.
• Association with TME and Genomics:
  • Low-TRS Group: Characterized by abundant immune cell infiltration, higher expression of MHC molecules, chemokines, and immune checkpoints (PD-1, PD-L1, CTLA-4), and elevated activity across all seven steps of the cancer immune cycle.
  • High-TRS Group: Associated with higher tumor purity, lower immune infiltration, and higher Tumor Mutation Burden (TMB). Notably, the high-TRS group showed a higher frequency of TP53 mutations (49% vs. 20% in low-TRS) and a higher TIDE score, indicating immune dysfunction and exclusion.
• Immunotherapy Prediction: In three independent immunotherapy cohorts, patients with low TRS scores exhibited significantly better overall survival and higher response rates (Complete/Partial Remission) to PD-1/PD-L1 blockade compared to high-TRS patients. Low TRS was strongly associated with the "immunoinflammatory" (inflamed) tumor phenotype.

Significance and Claims
The authors claim that this study successfully developed a robust, machine learning-derived T-cell-associated signature (TRS) that serves as an independent prognostic factor for breast cancer. The significance of the work lies in:

1. Superior Predictive Power: The TRS outperforms traditional clinical staging and molecular subtyping in predicting patient survival.
2. Immunotherapy Guidance: The signature effectively stratifies patients likely to benefit from immune checkpoint inhibitors, identifying a "low-TRS" population with a "hot" tumor microenvironment and favorable response to immunotherapy.
3. Mechanistic Insight: The study links specific T-cell-associated gene expression patterns to immune infiltration, TMB, and immune checkpoint expression, providing a biological basis for the observed prognostic differences.

Limitations
The authors acknowledge several limitations:

• Retrospective Nature: All data were derived from public, retrospective cohorts, potentially introducing selection bias.
• Sample Size: Some immunotherapy validation cohorts had relatively small sample sizes (e.g., n=27, n=28).
• Mechanistic Validation: The specific biological functions of the genes included in the signature have not been experimentally validated via in vitro or in vivo models.
• Lack of Prospective Data: The model has not yet been validated in prospective clinical trials.

Conclusion
The study concludes that the TRS model is a powerful tool for personalized prognostic prediction and clinical decision-making in breast cancer, particularly for identifying patients who may benefit from immunotherapy. Future work is required to validate these findings in prospective multicenter studies and to elucidate the underlying molecular mechanisms.