MolSculptor: a diffusion-evolution framework for multi-site inhibitor design

Complex diseases, such as cancer and neurodegenerative disorders, often involve the coordinated regulation of multiple binding sites. An emerging strategy is the design of multi-site inhibitors, which achieve affinity and selectivity for multiple binding pockets through a single molecule. Current molecular design methods based on deep generative models (DGMs) typically rely on suitable protein-ligand training datasets. However, datasets for multi-site inhibitors are limited, constraining the applicability of DGMs in this domain. In this work, we introduce MolSculptor, a training-free framework for multi-site inhibitor design. By combining a latent diffusion model with an evolutionary algorithm, MolSculptor enables the multi-objective de novo design and molecular optimization of multi-site inhibitors without training procedure. Furthermore, we applied MolSculptor to various real-world tasks, including three dual-target inhibitor optimization tasks and a selective inhibitor design task. The results demonstrated that our framework enables efficient design of multi-site inhibitors, outperforming state-of-the-art methods.