Sesame Generates Molecules with Structure-Aware Spatial Density-Map Conditioning.

Generative molecular models are a promising frontier in drug design, but they face significant challenges in understanding small molecule structure, protein-ligand interactions, and generating molecules from scratch. A critical, yet often overlooked, capability is the ability to grow a molecule from a partial starting point, such as a scaffold or fragment provided by a chemist, which is essential for lead optimization. Researchers have developed Sesame (Spatial Evoformer for a Structure-Aware Molecular Engine), a diffusion-based molecular generation model designed to address these needs. Sesame incorporates a novel spatial pairformer module that conditions its generation process on both partial molecular structure and the surrounding protein pocket. This conditioning is achieved through continuous spatial density maps. This single, versatile conditioning mechanism supports both de novo molecule generation and fragment-conditioned lead optimization. It allows medicinal chemists to prune a hit compound to a scaffold and then have Sesame intelligently grow it in productive ways. The model also features a diffusion framework for jointly denoising atom types, bond types, and positions, along with a trajectory finetuning scheme to enhance generation quality, trained on extensive ligand and protein-ligand datasets.