Pharmacogenomic impact and genetic architecture of toxicity in pediatric acute lymphoblastic leukemia induction therapy: an exploratory modeling approach

Induction chemotherapy for pediatric Acute Lymphoblastic Leukemia (ALL) is highly curative but accompanied by a severe toxicity profile. Evaluating the genetic susceptibility to these adverse events could inform hypothesis-generating risk stratification. In this exploratory retrospective cohort study (N = 178), we investigated the influence of 14 candidate pharmacogenetic variants on the development of prevalent Adverse Drug Reactions (ADRs, ≥ 10 events). Multiple logistic regression, adjusted for basic demographic and anthropometric covariates (age and body surface area), was utilized across comprehensive inheritance architectures (additive, dominant, recessive, and co-dominant). For clinical translation hypothesis generation, we applied LASSO-penalized logistic regression (α = 1), 10-fold cross-validation to derive parsimonious Pharmacogenetic Risk Scores (PGRS). Model performance was evaluated using the optimism-corrected Area Under the Receiver Operating Characteristic Curve (AUC) via 1000-iteration bootstrap resampling, and calibration was assessed via the Brier score. We observed several complex interactions, notably suggesting an increased risk of severe neutropenia in patients carrying a homozygous variant CYP3A5 rs776746 genotype (nominally significant OR 5.58, p = 0.006) and a potential pleiotropic protective effect from NFATC2 rs6021191 (OR 0.09, p = 0.001). Penalized selection allowed the construction of compact, multi-locus mathematical risk scores that demonstrated promising diagnostic discrimination for Neutropenia (optimism-corrected AUC 0.793, 95% CI 0.72–0.85; Brier score 0.12), while displaying limited to moderate performance for other clinically discriminating toxicities like Infection (AUC 0.650) and Gastrointestinal complications (AUC 0.607). The primary contribution of this study is the prioritization of pharmacogenomic candidate signals rather than the proposal of models ready for clinical intervention. The integration of alternative genotype parameterizations and regularized machine learning highlights pathways for future predictive research. However, given the retrospective nature of the study and inherent risks of internal optimism, these scores serve as biologically informed proof-of-concept models that strictly require robust prospective external validation.