Computational & AI Journal of Chemical Information and Modeling
A graph-neural-network screening pipeline, backed by physics-based rescoring and lab assays, repurposes the beta-blocker carvedilol as a partial THRβ agonist
- A tiered workflow first used a graph neural network to rank compounds from an internal library, then applied docking and MM/GBSA rescoring to narrow the list to carvedilol.
- The network model beat conventional machine-learning baselines in retrospective comparison.
- Biochemical TR-FRET and cell-based experiments showed partial agonism with coactivator preference similar to resmetirom, plus reduced lipid load and induction of metabolic genes.
- Molecular dynamics suggested stabilization in the pocket by hydrophobic contacts and two key hydrogen bonds.
Carvedilol is a promiscuous, well-characterized drug, so ask whether the GNN actually found it or whether docking and MM/GBSA did the selecting; an ablation without the network would tell you. "Beat baselines" retrospectively says little unless splits were scaffold-based. Partial agonism and coactivator bias are interesting, but what you'd check next is THRα versus THRβ selectivity and whether cell activity occurs at clinically reachable exposures.
Paper: Journal of Chemical Information and Modeling ↗ AI-Driven Discovery Pipeline Identifies Carvedilol as a Novel Partial Agonist of THRβ