Sesame: Structure-Aware Molecular Generation via Spatial Density-Map Conditioning

How can a single generative model effectively handle both de novo molecular design and lead optimization—the process of refining a known chemical hit—while maintaining structural awareness of the target protein pocket?

The authors introduce Sesame (Spatial Evoformer for a Structure-Aware Molecular Engine), a diffusion-based model that represents both protein pockets and partial molecular fragments as continuous spatial density maps. By using a novel spatial pairformer module, the model conditions its generation process on these density fields. This allows the system to treat a scaffold or fragment as a soft prior, enabling the model to grow and elaborate the structure within a protein pocket. The architecture also features a hybrid discrete-continuous diffusion process to jointly denoise atom types, bond types, and 3D coordinates, alongside a trajectory finetuning scheme that improves generation quality by training on the model's own sampling rollouts.

Sesame successfully integrates de novo design and lead optimization into a single conditioning framework. Because both the protein pocket and the starting fragment are represented as density maps, the model requires no architectural changes to switch between these tasks. In fragment-conditioned generation, the model demonstrates high fidelity to the input, with 94.8% of generated molecules retaining the seeding fragment as a substructure. The use of a pairformer architecture, inspired by protein-folding models, allows the system to effectively capture complex multi-way interactions between the ligand and the protein pocket, resulting in chemically sensible and pocket-compatible molecular structures.

This work bridges the gap between generic molecular generation and practical medicinal chemistry. By allowing chemists to prune a hit molecule to a scaffold and having the model grow it, Sesame provides a tool that complements human expertise rather than replacing it. The unified representation of structural priors as density maps offers a flexible, efficient way to perform structure-based drug design without the rigid constraints found in many existing generative models.