Key Takeaways:
Tessel Bio released Sesame, a diffusion-based generative chemistry model that designs drug-like molecules to fit any protein pocket, either from scratch or by growing a chemist's scaffold.
A new spatial pairformer module reads both a partial molecule and the surrounding protein pocket as continuous spatial density maps, handling de novo design and lead optimization in one system.
Tessel published Sesame openly in a preprint on arXiv, adding to a growing wave of AI tools built to speed early drug discovery.
Tessel Bio released Sesame on June 26, a generative chemistry model that designs drug-like molecules to fill any protein pocket, the company laid out in a preprint posted to arXiv. The model works two ways, building a molecule from nothing or growing one from a fragment a chemist supplies.
Sesame, short for Spatial Evoformer for a Structure-Aware Molecular Engine, is built on diffusion, the same class of method behind modern image generators. Its core addition is a spatial pairformer module that reads a partial molecule and the shape of the protein pocket around it as continuous density maps, then fills in atom types, bond types, and positions together.
That single mechanism covers two jobs that usually need separate tools. A medicinal chemist can prune a promising hit down to a scaffold and have Sesame grow it in new directions, which is the central move in lead optimization. The authors also describe a training step that learns from the model's own sampling runs to sharpen the molecules it proposes.
Tessel framed the release as open and unrestricted. Naren Tallapragada of Tessel said the model fills any pocket with "no constraints on size or class," a claim aimed at chemists who have watched earlier generative models handle only narrow cases.
Generative chemistry has been one of the busiest corners of AI for drug design, and a model that does de novo design and optimization in one pass lowers the bar for smaller labs to test it. The release lands alongside other efforts to point machine learning at biology and medicine, from a $500 million push to build open AI-trainable datasets of human cells to an AI company building a body scanner it says undercuts the MRI. Whether Sesame's molecules hold up in the lab is the test that matters, and that evidence comes slower than a preprint.
People Also Ask
What is Sesame, the generative chemistry model from Tessel Bio?
Sesame is a diffusion-based AI model that designs drug-like molecules to fit a target protein pocket, either from scratch or by extending a partial molecule supplied by a chemist. Tessel Bio described it in a June 2026 arXiv preprint.
How does Sesame design a molecule for a protein pocket?
It uses a spatial pairformer module that reads the partial molecule and the surrounding pocket as continuous spatial density maps, then jointly generates atom types, bond types, and positions through a diffusion process.
What is the difference between de novo generation and lead optimization?
De novo generation builds a molecule from nothing, while lead optimization grows or refines an existing hit, usually from a scaffold a chemist selects. Sesame handles both with one conditioning mechanism.
Is Sesame available to researchers?
Tessel Bio released Sesame openly and published a preprint on arXiv that lays out the model's architecture and training approach.
Sources: Sesame preprint on arXiv (Konstantin Yatsenko and Arvind Thiagarajan). Naren Tallapragada on X. Tessel Bio.