MolWorld: Molecule World Models for Actionable Molecular Optimization

Researchers introduce MolWorld, a novel AI framework that optimizes molecular structures for drug discovery by modeling actionable pathways between molecules. Unlike existing methods, MolWorld ensures discovered candidates are chemically reachable from known compounds through valid intermediate steps, making them practically viable for lead optimization.

MolWorld addresses a critical gap in computational drug discovery: the difference between theoretically optimal molecules and practically achievable ones. Traditional de novo optimization methods generate high-scoring candidates without ensuring they can be synthesized or reached through realistic chemical transformations from existing compounds. This disconnect between prediction and feasibility has limited the real-world applicability of AI-driven molecular design. MolWorld reframes the problem as sequential graph expansion, where each molecule connects to others through valid local transformations, creating a traversable chemical space that mirrors how medicinal chemists actually work—iteratively modifying known scaffolds rather than inventing entirely new structures. The framework uses a learned world model to evaluate candidates and update the molecule-transfer graph, maintaining structural connectivity throughout optimization. This approach bridges computational chemistry and practical synthesis workflows, reducing the gap between computational predictions and laboratory implementation. The experimental results demonstrate that MolWorld discovers molecules with strong target properties while preserving the interpretability and accessibility that chemists require for lead series evolution. For the pharmaceutical and biotech industries, this represents meaningful progress toward deployable AI tools that enhance rather than replace expert chemists' decision-making. The methodology's emphasis on actionability and sequential design could accelerate drug discovery timelines by eliminating infeasible candidates early and providing chemically sensible optimization pathways. As AI increasingly influences drug development pipelines, frameworks that respect both computational performance and practical chemistry constraints will likely become essential infrastructure.

• →MolWorld ensures optimized molecules are reachable from known compounds through valid chemical transformations, addressing a major limitation of existing de novo methods.
• →The framework models molecular optimization as sequential graph expansion with intermediate steps, mirroring how medicinal chemists actually conduct lead optimization.
• →Experimental results show MolWorld achieves high property optimization while maintaining substantially stronger structural connectivity than baseline approaches.
• →The approach bridges the gap between computational predictions and laboratory feasibility, making AI-generated candidates more actionable for drug discovery workflows.
• →This development could accelerate pharmaceutical development by combining computational rigor with practical synthetic chemistry constraints.