Development and Validation of an Anoikis-Related Machine Learning Signature for Prognosis and Brain...

Background: Brain metastasis is associated with poor prognosis in lung adenocarcinoma (LUAD). Anoikis resistance may contribute to tumor cell survival during metastatic dissemination and brain colonization; however, robust biomarkers for prognostic stratification and brain metastasis-associated classification remain limited. This study aimed to investigate anoikis-related molecular features in LUAD brain metastasis and develop a machine learning-based signature for prognostic assessment and exploratory classification of primary and brain-metastatic LUAD samples. Methods: We integrated single-cell and multi-cohort bulk transcriptomic data. Single-cell analysis was performed to characterize anoikis-related cellular states and intercellular communication in primary and brain-metastatic LUAD samples. In the bulk transcriptomic analysis, TCGA-LUAD was used for prognostic feature selection and risk-model construction, and GSE26939 was used for external prognostic validation. The classification performance of the fixed signature for distinguishing primary LUAD from brain-metastatic LUAD samples was further evaluated in GSE161116 and GSE271259. Immune microenvironment features were assessed, and an LLM-assisted exploratory drug-screening strategy combined with molecular docking was used to prioritize candidate compounds. Results: Single-cell analysis suggested that metastatic epithelial cells exhibited enhanced anoikis-related activity, accompanied by macrophage-associated SPP1-CD44 and MIF-(CD74+CXCR4) communication patterns. Machine learning-based feature selection identified an eight-gene signature consisting of BIRC3, CCL20, CLEC7A, CTSL, GOLM1, ICAM3, MTUS1, and SERPINH1. The signature showed prognostic value in TCGA-LUAD and GSE26939 and demonstrated exploratory classification performance in distinguishing primary LUAD from brain-metastatic LUAD samples. High-risk patients exhibited immune microenvironment alterations and enrichment of tumor progression-related pathways. LLM-assisted compound prioritization and molecular docking highlighted resveratrol and SB431542 as hypothesis-generating candidates with predicted interactions with core targets. Conclusions: This study identified an anoikis-related eight-gene signature for LUAD prognostic stratification and exploratory brain metastasis-associated classification. The findings suggest the potential involvement of anoikis-related tumor-microenvironment interactions in LUAD brain metastasis and provide candidate genes and compounds for further experimental validation.