Machine learning prediction for AML based on 3D genome selected circRNA

Acute myeloid leukemia (AML) is a clinically aggressive hematologic malignancy driven by complex genetic and epigenetic aberrations. Circular RNAs (circRNAs), characterized by covalently closed structures and exceptional stability, have emerged as promising diagnostic biomarkers. However, existing circRNA-based predictive models largely depend on differential expression, overlooking the potential impact of higher-order chromatin organization on circRNA formation and function. Here, we propose a machine learning framework that integrates three-dimensional (3D) genome architecture to refine circRNA selection for AML prediction. By mapping 9,565 circRNAs onto a 3D chromatin model reconstructed from Hi-C data, we analyzed their spatial clustering and biological pathway enrichment. Eighteen pathways exhibited significant 3D aggregation of circRNAs, enabling radial stratification based on nuclear localization. Five circRNA panels were designed using complementary strategies combining expression, pathway, and spatial features. Cross-validation and external validation across six machine learning algorithms showed that the panel derived from the fifth radial layer (Panel-3DG-Radius5) achieved the most robust and consistent performance (ROC-AUC > 0.99). Integrating 3D genomic context reduced feature collinearity while enhancing biological interpretability. Overall, our study establishes a 3D genome-informed paradigm for circRNA biomarker discovery, demonstrating that spatial genome organization can substantially improve the precision and robustness of AML predictive modeling.