MolWorld: Molecule World Models for Actionable Molecular Optimization

Researchers have introduced MolWorld, a novel framework designed to enhance molecular optimization in drug discovery by ensuring candidates are not only high-performing but also actionable. Traditional methods often overlook the practical reachability of new molecules from known compounds. MolWorld addresses this by modeling optimization as the sequential expansion of a molecule-transfer graph, where nodes represent molecules and edges denote valid local structural transformations. The system employs a learned world model to guide the search process, iteratively selecting anchor contexts, generating conditioned candidate molecules, and evaluating their properties. Admissible candidates are retained and integrated into the evolving graph, which subsequently directs further optimization steps. Experimental results on property optimization and docking tasks demonstrate that MolWorld successfully identifies molecules with improved properties while maintaining robust structural connectivity to known compounds. This approach supports more plausible and sequential molecular design, bridging the gap between theoretical predictions and practical synthetic feasibility in pharmaceutical development.