Bayesian Additive Regression Trees modeling of pharmacogenomic and clinical data predicts amlodipine response in hypertension

BACKGROUND: The adoption of precision medicine in antihypertensive treatment remains challenging, largely because of substantial interindividual variability in drug response. This study aimed to determine whether genotyping combined with Bayesian machine-learning methods could improve the prediction of response to amlodipine. METHODS: We conducted a multicenter prospective cohort study of 292 Chinese patients with hypertension treated with amlodipine. Eight single-nucleotide polymorphisms (SNPs) in key pharmacokinetic and pharmacodynamic pathways were sequenced. After excluding two SNPs with insufficient variation, six SNPs were integrated with clinical variables to develop prediction models. The outcome, blood pressure control as a measure of treatment efficacy, was defined as achieving office SBP less than 140 mmHg and DBP less than 90 mmHg after treatment. Using a rigorous validation framework, we compared Bayesian Additive Regression Trees (BART) with Random Forest Classification (RFC) and Bayesian linear modeling through repeated cross-validation and external validation. RESULTS: At the rs5065 locus, 217 individuals (99.5%) carried the wild-type genotype, and rs2740574 was exclusively homozygous for the mutant allele. Therefore, only the remaining six SNPs were used for model development, comparison, and evaluation. The BART model achieved an external validation area under the receiver operating characteristic curve (AUC) value of 0.846, notably outperforming RFC (AUC = 0.61) and Bayesian linear models (AUC = 0.644-0.665). CONCLUSION: BART substantially improves prediction of blood pressure control in patients with hypertension treated with amlodipine monotherapy. Integrating pharmacogenomic data with advanced Bayesian machine learning offers a robust framework for personalized antihypertensive therapy.