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

Integration of six polymorphisms in amlodipine pharmacokinetic and pharmacodynamic pathways with clinical variables in a non-linear Bayesian model

Interindividual variability in antihypertensive response remains the main obstacle to personalized prescribing. This multicenter prospective cohort included 292 Chinese patients with hypertension treated with amlodipine, in whom eight polymorphisms in pharmacokinetic and pharmacodynamic pathways were sequenced. Two were excluded for insufficient variation — 217 subjects (99.5%) carried the wild-type genotype at rs5065 and rs2740574 was exclusively homozygous for the mutant allele — and the remaining six were combined with clinical variables. The outcome was blood pressure control, defined as office systolic blood pressure below 140 mmHg and diastolic below 90 mmHg after treatment. Using repeated cross-validation followed by external validation, the Bayesian Additive Regression Trees model reached an area under the curve of 0.846, versus 0.61 for random forest classification and 0.644-0.665 for Bayesian linear models.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

A gap from 0.846 to 0.61 in area under the curve between two models trained on the same 292 patients and the same six variants is unusually wide and calls for caution, all the more as the external validation sample size is not stated in the abstract. The genetic contribution itself is not isolated: nothing shows what the model gains over clinical variables alone, which is the pharmacogenetic question. To be read as a methodological demonstration, not as a treatment decision aid.

Analysis by Dr Thibaut Benquey