MERIT predicts trial outcomes from mechanism alone, without compound-identity shortcuts, and locks outcome-blind phase III calls
- Many trial-outcome models lean on trial design, development history or compound identity and so memorize compounds; MERIT uses only molecular and disease features, linking intended and off-target effects through drug–protein, protein–metabolite and immune interaction maps to tissue-specific efficacy and safety.
- On a structure-holdout split over hundreds of small-molecule drugs and thousands of trials, it beat two published methods retrained on the same folds.
- It ranked the approved indication above the failed one for most of a preselected set of drugs that failed one indication but are approved for another, and the authors registered locked, outcome-blind predictions for ongoing phase III drug–indication pairs.
💡 Points Based on the abstract
Dropping compound identity and trial design is the right fix for leakage, and a structure-holdout split is a harder test than most. But an overall AUROC is not the question a portfolio team asks; the failed-here, approved-there ranking only matters if it holds on non-obvious cases, so I'd read those 60 compounds one by one. The locked phase III predictions are the real test, and I'd wait for those readouts before this goes near a go/no-go discussion.
MERIT: Mechanism-Driven Model Predicts Drug Outcomes and Nominates Indications for Failed Drugs ↗