Abstract
Acute myeloid leukemia (AML) exhibits significant heterogeneity in disease progression and therapeutic response, highlighting the urgent need for novel biomarkers to improve risk stratification and therapeutic targeting. In this study, we integrated multi-omics data from The Cancer Genome Atlas (TCGA, n = 151) and Genotype-Tissue Expression (GTEx, n = 337) cohorts to systematically analyze dynamic expression patterns of exosome-related genes in AML. Using univariate Cox regression and least absolute shrinkage and selection operator (LASSO) algorithms, we identified 13 exosome-associated genes (EXOSC4, TMEM109, THBS1, MYH9, HLA-DRA, CAPZB, ITGA4, MYL6, CYB5R1, PSMA2, MPO, NDST2, and CANX) and constructed a prognostic risk model. The model demonstrated superior predictive accuracy compared to traditional clinical parameters, with area under the curve (AUC) values of 0.819, 0.825, and 0.832 for 1-, 2-, and 3-year survival predictions in the training set, and 0.909 in the independent GEO validation cohort (GSE71014). Kaplan-Meier analysis revealed significantly shorter overall survival in the high-risk group (log-rank P < 0.001, hazard ratio = 0.22, 95% CI = 0.13-0.36). Immune microenvironment characterization using CIBERSORTx identified increased infiltration of regulatory T cells (Tregs, P < 0.01) in high-risk patients. Functional enrichment analysis revealed enrichment of PI3K-Akt signaling pathways and TP53 transcriptional networks in high-risk groups. Molecular docking studies confirmed strong binding affinity of verteporfin (ITGA4 inhibitor, docking score=-16.0 kcal/mol) and ebselen (MPO inhibitor) to their respective targets, suggesting potential therapeutic strategies to overcome chemotherapy resistance. This study establishes a robust 13-gene exosome-based prognostic signature for AML risk stratification and identifies novel immunomodulatory mechanisms mediated by exosome-driven Treg polarization.