Diffusion-Based Generative Model With Scaffold-Hopping Strategy Yields Highly Potent Bioactive Molecules

As a critical step in drug discovery, lead optimization is a profoundly complex endeavor with a notoriously high failure rate, as it necessitates the simultaneous optimization of multiple, often conflicting parameters, including physicochemical properties, drug-likeness, synthetic accessibility, and target binding affinity. While several generative models have been proposed for lead optimization under multi-property constraints, they still struggle to balance multi-objective optimization with sufficient scaffold-level exploration. To address this challenge, we present SMarT-Diff (Scaffold-based Multi-property Tuning Diffusion), a generative diffusion model that achieves this balance by reinventing scaffold hopping—enabling both property optimization and structural novelty. SMarT-Diff achieved superior performance across diverse molecular generation and optimization metrics. Notably, across both single-target (LRRK2, HPK1, GLP-1R) and dual-target (GSK3β/JNK3) molecular optimization tasks, the model consistently generated drug-like molecules exhibiting enhanced structural diversity, preserved pharmacophoric features, and high synthetic accessibility. Furthermore, wet-lab validation of our model-generated compounds against LRRK2 identified a highly promising candidate with an IC₅₀ of 1.544 nM, which surpasses even the positive control LRRK2-IN-1. This result not only confirms the compound's exceptional potency but also demonstrates the strong real-world potential of our model to drive the design and optimization of novel, highly effective drug candidates.