Cerebrospinal fluid proteomics to identify biomarkers for the diagnosis and prediction of leptomeningeal metastasis from non–small cell lung cancer.

Abstract

e15009 Background: Leptomeningeal metastasis (LM), a terminal complication of non-small-cell lung cancer (NSCLC), lacks validated diagnostic criteria. This study developed machine-learning classifiers that integrate cerebrospinal fluid (CSF) proteomic, clinical-epidemiologic and biochemical variables to enable accurate and minimally invasive diagnosis and prognostication of LM. Methods: CSF samples were collected from NSCLC patients undergoing lumbar puncture at the Second Affiliated Hospital of Nanchang University between September 2023 and December 2025 , followed by quantitative proteomic profiling. Integrative machine-learning models were trained using the proteomic and corresponding clinical datasets to identify optimal diagnostic biomarkers for LM. Parallel reaction monitoring (PRM) validation was subsequently performed in patients from the prospective cohort. Differential CSF proteome profiling and prognostic model construction were conducted in EGFR-mutant NSCLC patients with LM versus Non-LM controls. Results: 1350 proteins differed between LM (n = 61) and Non-LM (n = 34) in the retrospective cohort. Machine model selecting a minimal 4-protein panel (C7, FZD9, IL1RAP, SERPINA3) that distinguished LM from Non-LM with AUC 0.934, sensitivity 86.5%, specificity 95.0%. Parallel reaction monitoring (PRM) confirmed quantitative differences. In EGFR-mutant LM, down-regulation of FKRP, ERBB4, and MACF1 together with up-regulation of ASS1 predicted longer overall survival. Conclusions: CSF proteome-clinical machine-learning models provide a robust, minimally-invasive 4-protein diagnostic signature for NSCLC-LM and a separate survival classifier for EGFR-mutant disease, supporting early detection and risk-stratified management.