Abstract
Despite significant advancements in overall survival rates for childhood acute lymphoblastic leukemia (ALL), relapse continues to pose a major challenge. MicroRNAs have proven valuable for improving diagnosis, treatment, and survival outcomes, establishing themselves as key biomarkers. Using RNA-seq data from 123 ALL patients and employing predictive modeling via automated machine learning (AutoML) alongside causal-inspired biomarker discovery, we identified highly predictive microRNA signatures linked to high-risk strata and clinical features in unfavorable cases. We further identified predictive signatures for each genetic subtype of childhood ALL, highlighting shared miRNAs throughout the study. A thorough literature review of the relationships between miRNA differential expression and key high-risk features in childhood ALL [immunophenotype, elevated white blood cell counts at diagnosis, central nervous system involvement, measurable residual disease (MRD), and chemoresistance] confirmed the signatures generated in this study. Our results revealed a highly predictive signature distinguishing B- and T-ALL, associated with apoptosis, confirming the reported difference between the two immunophenotypes. Additionally, miR-223 emerged as crucial for high-risk stratification and chemoresistant MRD-positive cases. These findings demonstrate the potential of AutoML tools to reveal novel biological insights in pediatric ALL, driving future advancements.