Abstract
Abstract
Molecular descriptors are frequently used to interpret quantitative structure–activity relationship models and to guide lead optimization. However, translating a specified descriptor change into concrete chemical structures remains challenging. Here, we developed Scaffold-tuner, a property-directed compound editing tool that proposes scaffold-preserving structural modifications toward specified changes in hydrogen-bond donors, hydrogen-bond acceptors, aromatic rings, and rotatable bonds. Scaffold-tuner was evaluated using 1,000 drug-like compounds from ChEMBL and compared with random local editing, chemically reasonable mutations (CReM), and REINVENT4 Mol2Mol. Scaffold-tuner achieved the intended descriptor change in seven of eight evaluated tasks and showed particularly high efficiency when only a small number of candidates were considered. With one evaluated candidate, the mean parent success rate was 42.3%, compared with 13.9%, 15.6%, and 17.1% for random editing, CReM, and REINVENT4, respectively. Although broader generative approaches achieved higher coverage with larger candidate budgets, Scaffold-tuner retained relatively high similarity to parent structures. These results suggest that Scaffold-tuner provides a useful framework for translating specified descriptor changes into a limited number of interpretable, parent-like candidate structures.