For topiramate in alcohol use disorder, drinking history still beats genes

Design

Machine-learning (random forest) analysis of randomised trial data

Population

278 participants with alcohol use disorder treated with topiramate

Primary outcome

Predicted topiramate response from clinical and genetic features

Effect

Parsimonious model R-squared 0.30; drinking measures strongest, time-to-relapse polygenic score secondary

Topiramate helps some people with alcohol use disorder but response varies, and the hope is that genetics could pick responders. This analysis applied machine learning to a randomised trial of 278 participants, testing 23 pre-treatment features including four alcohol polygenic risk scores.

The result is a useful corrective to pharmacogenomic enthusiasm. The strongest predictors were not genetic but behavioural: average drinks per day and percentage of heavy drinking days. A polygenic score for time until relapse did add meaningfully, and the parsimonious three-predictor model reached a bias-corrected R-squared of 0.30. Likely responders had lower baseline drinking severity and a higher time-to-relapse polygenic score.

The practical message is that a careful drinking history remains the best guide to who will respond to topiramate, and that pharmacogenomic scores are a promising add-on rather than a test to order today. Precision prescribing for alcohol use disorder is advancing, but behavioural data still carry most of the signal.

• Machine-learning analysis of a topiramate trial in alcohol use disorder, 278 participants.
• Average drinks per day and percentage of heavy drinking days were the strongest predictors.
• A polygenic score for time until relapse added meaningfully but was secondary.
• The parsimonious three-predictor model reached an R-squared of 0.30.
• A careful drinking history remains the best current guide to topiramate response.

Why it matters

It tempers the assumption that genetics can yet pick topiramate responders, keeping the focus on clinical data.

Don't overread it

This is exploratory modelling with an R-squared of 0.30; the polygenic score was a secondary contributor and is not ready to select patients in clinic.

The statistics, in plain English

An R-squared of 0.30 means the model explains under a third of the variation in response, useful for research but far from deterministic for an individual. That drinking measures outweighed four polygenic scores is the key finding.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.