Predicting Prognosis of Locoregionally Advanced Nasopharyngeal Carcinoma Using Machine Learning Models Based on Plasma Proteomics : A retrospectively registered Study

Yuyi Li, Chao Tan, Xiaoyu Chen, Weichang Zhu, Cuihong Jiang, Lili He, Shuai Xiao, Changgen Fan, Xu Ye, Qi Zhao, Wenqiong Wu, Yanxian Li, Yanfang Qiu, Kailin Chen, Shulu Hu, Pan Chen, Feng Liu, Hui Wang

Abstract Background Patients with locoregionally advanced nasopharyngeal carcinoma (LA-NPC) exhibit heterogeneous short-term responses despite induction chemotherapy plus concurrent chemoradiotherapy, and effective plasma protein prognostic markers are lacking. This study aimed to screen and validate differentially expressed proteins (DEPs) associated with short-term response using plasma proteomics and machine learning, to support personalized treatment. Methods 107 newly diagnosed LA-NPC patients (stage III-IVA) from Hunan Cancer Hospital were enrolled (October 2023-June 2025) and divided into a discovery cohort (n = 6) and a validation cohort (n = 101). Plasma samples were collected before treatment, after induction chemotherapy, and after radiotherapy. Data-independent acquisition (DIA) mass spectrometry-based proteomics was performed. LASSO regression and random forest were employed to identify key DEPs. Candidate proteins were validated by ELISA in the validation cohort. Predictive performance was assessed by ROC curve, multivariate ordinal logistic regression, Kaplan-Meier survival analysis, and multivariate Cox regression. Results Proteomics identified 67 DEPs between patients with complete response (CR) and stable disease (SD) before treatment, enriched in endoplasmic reticulum protein processing, ribosome, and neurodegenerative disease pathways. Intersection of LASSO (4 DEPs) and random forest (139 DEPs) yielded two key DEPs: BAK1 and GAA. ELISA confirmed that pre-treatment BAK1 was significantly higher in CR and partial response (PR) groups compared to SD (P < 0.05); post-treatment, BAK1 increased in SD but decreased in CR/PR groups. GAA showed no significant differences. Pre-treatment BAK1 predicted short-term response with an AUC of 0.902. Multivariate ordinal logistic regression showed BAK1 was significantly associated with response stratification (P = 0.039). Based on the optimal cut-off, the BAK1 high-expression group had significantly better distant metastasis-free survival (DMFS) and progression-free survival (PFS) (P < 0.05), and BAK1 was an independent factor for both. Conclusion This study is the first to discover and validate plasma BAK1 as a novel predictive biomarker for short-term response to induction chemotherapy plus concurrent chemoradiotherapy in LA-NPC, with additional prognostic value for DMFS and PFS. BAK1 holds promise for early risk assessment and individualized treatment strategies.