Integrating inheritance model selection, weighted polygenic scoring, and interpretable machine learning for cardiotoxicity risk prediction in pediatric ALL

Abstract Background Anthracycline-induced cardiotoxicity adversely affects treatment outcomes in acute lymphoblastic leukemia (ALL). Genetic variation in drug-metabolizing enzymes and membrane transporters may modulate individual susceptibility. This study investigated genetic predisposition to cardiotoxicity in ALL patients and evaluated cumulative risk using polygenic risk scoring and machine learning. Methods A total of 216 ALL patients receiving anthracycline-based induction therapy were enrolled. Cardiotoxicity was defined as a decrease in LVEF of more than 10% points to below 53%, or the presence of clinical cardiac abnormality after induction. Five variants (SOD2 rs4880, CBR1 rs9024, PNPLA3 rs738409, ABCC1 rs4148350, and ABCG2 rs2231142) were genotyped. Results ABCG2 rs2231142 was the only variant significant after Bonferroni correction, with the AC genotype conferring the highest risk under the codominant model (OR = 5.25, 95% CI: 1.84–14.95; p < 0.001; AIC = 240.70). CBR1 rs9024 showed a significant risk-increasing effect under the recessive model (OR = 1.70; p = 0.004). A beta-weighted PRS outperformed an unweighted additive score (AUC: 0.802 vs. 0.498), and tertile stratification demonstrated a significant dose-response relationship with cardiotoxicity rates rising from 11.0% to 62.5% (trend p < 0.001). Among five machine learning classifiers, LightGBM without oversampling and Decision Tree with SMOTE both achieved AUC = 0.73. Conclusions ABCG2 rs2231142 is a robust single-variant determinant while CBR1 rs9024 emerged as a second independently significant risk variant under the recessive model in multivariable logistic regression. SOD2, PNPLA3, and ABCC1 variants, individually below significance thresholds contributed meaningfully when aggregated into a weighted PRS, underscoring the complementary value of polygenic over single-variant analyses. Integrating genetic predictors with machine learning offers a promising framework for personalized cardiotoxicity risk stratification in pediatric ALL.