Pathway-Informed Machine Learning Identifies Genetic Predictors of High-Dose Methotrexate-Induced Mucositis in Pediatric Acute Lymphoblastic Leukemia

Abstract

High-dose methotrexate for pediatric cancer treatment is frequently associated with mucositis, which can lead to delayed or discontinued treatment and impact survival. While individual genetic variants have been implicated, the cumulative impact of genetic variation within relevant biological pathways remains unexplored. We evaluated single nucleotide polymorphisms across 18 pathways previously identified as relevant to mucositis in 278 pediatric patients with acute lymphoblastic leukemia from six academic health centers across Canada. Pathway enrichment was assessed using the Joint Association of Genetic variants tool, and a predictive model was developed using XGBoost, a supervised machine learning algorithm based on gradient-boosted decision trees. Pathway enrichment identified significant associations in IL6 (P = 0.04) and WNT/β-catenin (P = 0.048) signaling pathways. The predictive model (area under the curve [AUC] = 0.76) highlighted single nucleotide polymorphisms associated with inflammation- and mucosa-related genes, including PRKCD, IL17B, MAST3, and CAPN9, with both risk and protective effects. Model performance dropped by 0.15 in AUC (from 0.76 to 0.61) after removing single nucleotide polymorphism features, underscoring their predictive value. This pathway-informed approach identifies genetic contributors to methotrexate-induced mucositis and supports polygenic risk prediction. Our findings provide a foundation for individualized toxicity risk profiling and suggest potential therapeutic targets to mitigate treatment-limiting mucositis in pediatric oncology.