MOOSE: Targeted Small Molecule Generation via Multi-Objective-Guided Discrete Diffusion

We introduce MOOSE, a Multi-Objective Optimization framework for property-driven Small-molecule Engineering using discrete diffusion. MOOSE builds on GenMol, a masked discrete diffusion model that generates molecules in a fragment-based token space, and adapts it to goal-directed molecular design through iterative fine-tuning toward desired property distributions. By scoring target binding affinity with a rapid, state-of-the-art sequence-based surrogate model (Bonbon), together with standard chemical heuristics for drug-likeness and synthetic accessibility, MOOSE enables scalable optimization without reliance on docking during generation. We evaluate MOOSE on a diverse set of protein targets relevant to therapeutic discovery. Across all targets, molecules generated by MOOSE achieve stronger predicted binding affinity compared to known reference ligands, while maintaining chemical diversity, synthesizability, and drug-likeness. Together, these results demonstrate that MOOSE can be directly integrated into practical small-molecule discovery pipelines, enabling scalable multi-objective optimization without docking-in-the-loop.