New Recipe Optimizes Molecular Design with Discrete Diffusion Models

Molecular optimization often begins with a pre-trained generative model that understands valid molecular structures. The challenge then shifts to adapting this model to efficiently discover molecules with specific, high-reward properties using a limited budget for oracle evaluations. This research systematically investigates various online adaptation choices for discrete diffusion models in this context. The study explores how different components—candidate acquisition, reward shaping, feedback reuse, and managing the shift from the prior—interact within a full online adaptation loop. Through controlled experiments across multiple small-molecule binding-affinity and protein-fitness tasks, the researchers found that acquisition, reward shaping, and model debiasing are complementary and crucial for achieving higher rewards, particularly for small molecules. Furthermore, incorporating replay mechanisms stabilizes learning, while validity penalties help maintain molecular integrity during exploration. The culmination of these findings is a practical recipe for feedback-efficient molecular optimization: online fine-tuning enhanced with acquisition, reward shaping, debiasing, replay, and validity control. This comprehensive approach consistently outperforms offline fine-tuning and inference-time search baselines, especially when the target high-reward molecules require significant deviation from the initial pre-trained distribution.