Construction and validation of a palmitoylation-related prognostic model for lung adenocarcinoma based on integrated bioinformatics and machine learning.

Background: Lung adenocarcinoma (LUAD) is a prevalent malignancy whose therapeutic management is complicated by nonspecific early symptoms, late-stage diagnosis, and aggressive metastasis. Given the critical role of palmitoylation in LUAD progression, this study aims to construct a prognostic model based on palmitoylation-related genes, identify key biomarkers, and elucidate their underlying mechanisms. Methods: In this study, bulk RNA-seq data and clinical information for LUAD were retrieved from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) databases, while palmitoylation-related genes were obtained from the GeneCards database. The LUAD gene expression profile was analyzed using DESeq2 and weighted gene co-expression network analysis (WGCNA) to identify potential palmitoylation-associated biomarkers. Subsequently, by exploring 117 combinations of 10 machine learning algorithms, a predictive prognostic model based on palmitoylation-related features was constructed. This model was trained on the TCGA cohort (n=493) with 10-fold cross-validation and externally validated using three independent LUAD datasets (GSE30219, GSE72094, and GSE31210). Features independently predictive of prognosis were identified by integrating baseline clinical characteristics and were used to construct a prognostic nomogram. The model's performance was rigorously evaluated through multi-faceted assessments, including the concordance index (C-index), Kaplan-Meier (KM) survival analysis, time-dependent receiver operating characteristic (ROC) curves, and comparative analysis with existing LUAD models from the past year. Core genes were further screened, and their expression patterns and prognostic significance were analyzed. Finally, patients were stratified into high- and low-risk groups based on LUAD palmitoylation-related genes (LPRGs), and differences in genomic alterations, immune microenvironment characteristics, and drug susceptibility between the groups were investigated. Results: Through differential expression analysis, 152 potential candidates were identified. Using machine learning, a prognostic signature comprising 51 LPRGs was constructed. The optimal model, "Random Survival Forest (RSF) + Ridge", demonstrated a C-index of up to 0.68 in the training set and achieved 0.70 in external validation cohorts. It outperformed other LUAD prognostic models published in the past year across both training and testing datasets. Univariate and multivariate Cox regression analyses confirmed the LPRGs signature and disease stage as independent prognostic predictors, which were subsequently incorporated into a clinical nomogram. A high-risk score was associated with poorer overall survival. Survival analysis indicated that elevated expression of TXN and DNAJB4 was linked to worse outcomes, whereas upregulation of SCN2B, GPD1L and ATP8A2 was correlated with favorable prognosis. Significant disparities were observed between the high- and low-risk groups regarding immune cell infiltration levels, immune functional activity, gene mutation frequency, and anticancer drug susceptibility. High-risk individuals exhibited increased mutation burden, reduced immune infiltration, and a weaker response to immunotherapy. In contrast, the low-risk group demonstrated enhanced drug sensitivity and lower tumor mutational burden. Conclusions: Our work developed a robust LPRGs-based prognostic model and nomogram that personalizes LUAD management and guides therapeutic decisions.