Abstract
INTRODUCTION: Primary membranous nephropathy (PMN) is one of the leading causes of nephrotic syndrome in adults, but reliable prognostic assessment tools remain limited. Although traditional clinical parameters and serum anti-PLA2R antibodies constitute cornerstones of prognosis assessment, they fail to fully capture the heterogeneity of PMN treatment responses. Identification of multidimensional prognostic risk factors facilitates further guidance for treatment in PMN patients. This study aimed to identify novel urinary protein signatures predictive of clinical remission and delineate molecular subtypes associated with relapse.. METHODS: We performed quantitative proteomic profiling on extracellular vesicles (uEVs) isolated from baseline urine samples of 86 biopsy-confirmed PMN patients using nanoflow high-performance liquid chromatography-tandem mass spectrometry (nanoHPLC-MS/MS). The primary endpoints were clinical remission (complete or partial remission) and time to clinical remission. A prognostic model was developed by screening 101 machine learning algorithms, with the final risk score derived from a random survival forest and stepwise Cox regression, internally validated via bootstrap resampling. Relapse-associated molecular subtypes were identified using nonnegative matrix factorization (NMF). RESULTS: During a median follow-up of 8 months (IQR: 3.4-18.5), 76.1% of patients achieved remission, with a median time to remission of 11.2 months (95% CI: 6.2-18.0). The four-protein risk model (PON1, ACTBL2, RDX, TPP1) effectively stratified patients into high- and low-risk groups (Harrell's C-index = 0.729). The combined model integrating this proteomic signature with clinical features (anti-PLA2R Ab, age, eGFR) demonstrated superior predictive performance (Harrell's C-index = 0.744) compared to the clinical-feature-only model (Harrell's C-index = 0.636). To improve clinical applicability, we developed a web-based interactive Shiny application for individualized risk prediction in PMN patients. Additionally, proteomic clustering identified three distinct molecular subtypes (PMN1-3), with subtype PMN2 emerging as an independent predictor of relapse (OR = 10.26, 95% CI: 1.68-81.97; p < 0.05). CONCLUSION: The combined model incorporating four noninvasive urinary proteomic signatures and clinical characteristics significantly improved performance of predicting clinical remission in PMN patients compared to the clinical-feature-only model. Furthermore, molecular subtyping of the urinary proteome can identify patients at high risk for relapse. These findings provide a foundation for integrating advanced proteomics into personalized prognostic assessment for PMN patients, pending external validation in larger cohorts.