Abstract / Summary
Introduction: Topiramate is efficacious for treating alcohol use disorder (AUD), but treatment response varies considerably. To advance precision treatment, we used machine-learning methods integrating clinical trial and genetic data to estimate expected response to topiramate.
Methods: Random forest (RF) regression models were used to estimate expected topiramate response using randomized clinical trial data from 278 participants and 23 pre-treatment features: four alcohol polygenic risk scores (PRS), nine alcohol drinking measures, and 10 clinical and sociodemographic variables. RF importance scores quantified each feature's contribution to predictive performance. Likely Responders (LRs) were identified using quintiles of predicted treatment response and counterfactual placebo response, with LRs defined as quintiles in which randomized trial data demonstrated topiramate superiority over placebo. Variable Selection Using Random Forests (VSURF) was used to eliminate less informative and redundant predictors.
Results: Two pre-treatment drinking measures consistently had the highest importance scores across all VSURF levels. The final parsimonious model retained three predictors: average drinks per day, percent heavy drinking days, and the PRS for "Time Until Relapse," with bias-corrected R2 = 0.30. LRs comprised the top four predicted response quintiles (80% of the sample) across models. Compared with unlikely responders, LRs had lower baseline drinking severity and higher "Time Until Relapse" PRS.
Conclusions: The "Time Until Relapse" PRS contributed meaningfully to estimating expected response, LR stratification, and observed topiramate treatment effects, although pre-treatment drinking measures remained the strongest predictors. These findings support incorporation of pharmacogenomic markers into machine-learning models to advance precision prescribing for AUD.