Leveraging single-cell trajectory analysis and machine learning to build a robust prognostic classifier for triple-negative breast cancer to pinpoint CCDC171 as a master regulator.

Abstract

1136 Background: Triple-negative breast cancer (TNBC) is limited by high heterogeneity and a lack of robust biomarkers for risk-stratified management. While chemotherapy is standard, traditional bulk signatures fail to capture high-resolution cellular transitions driving recurrence. This study integrates scRNA-seq with multi-algorithmic machine learning to identify key regulatory programs. We developed a trajectory-informed prognostic classifier to improve individual risk assessment and uncover potential therapeutic targets in the TNBC tumor microenvironment (TME). Methods: We performed single-cell RNA sequencing (scRNA-seq) on 24 TNBC samples to map the cellular landscape and applied pseudotime trajectory analysis to identify dynamic gene programs associated with cell-fate decisions. Prognostic genes derived from trajectory-informed differential expression were integrated with bulk RNA-seq data from TCGA-TNBC and validated in two independent cohorts (GSE58812, GSE135565). A robust prognostic model was constructed using an extensive machine-learning framework combining forward stepwise Cox regression and gradient boosting (StepCox[forward] + GBM). Model performance was evaluated using Harrell’s C-index and Kaplan–Meier analysis, and predictive utility was assessed in neoadjuvant treatment cohorts. Virtual gene knockout experiments were used to analyze the function of CCDC171. Results: We developed an 18-gene prognostic signature that robustly stratified TNBC patients into high- and low-risk groups across multiple cohorts (TCGA: HR = 55.12, p = 2.9×10 -12 ; GSE58812: HR = 3.96, p = 6.3×10 -4 ; GSE135565: HR = 7.73, p = 4.0×10 -3 ). The model outperformed existing prognostic signatures and generalized to non-TNBC breast cancer subtypes. Four genes (CRISP3, PDZK1IP1, CCDC171, IGFL1) were consistently retained across top-performing algorithms. CCDC171 emerged as a central regulator whose expression correlated with poor survival and coordinated a network involving complement activation (C1QB), metabolic reprogramming (CA8), and mitochondrial stress response (MTRNR2L8). In silico knockout of CCDC171 downregulated these effectors and suppressed oncogenic pathways including PI3K–Akt signaling, ECM interaction, and platinum resistance. Conclusions: Our study presents a TNBC prognostic model grounded in TME cellular dynamics and machine learning, with superior predictive accuracy and cross-subtype applicability. CCDC171 is identified as a novel master regulator of a multi-effector network driving tumor aggressiveness, offering a potential therapeutic target. This integrative framework supports risk-stratified management and personalized therapy in breast cancer.