A new model predicts how molecules interact with proteins
The Bonbon model identifies binding sites and mechanisms using paired sequences.
Researchers built Bonbon, a foundation model for molecular interactions. It works using approximately 3.2 trillion tokens of paired protein-ligand sequences.
The model was trained on protein-ligand sequences alone. It did not use structural coordinates or task-specific supervision.
Bonbon shows three zero-shot emergent capabilities. These include functional mechanism of action, residue binding site localization, and bond engagement.
Residue capabilities include orthosteric/allosteric discrimination. Bond capabilities distinguish between covalent and non-covalent engagement.
The model achieves state-of-the-art binding and affinity prediction. It does this at five orders of magnitude greater speed than co-folding methods.
In wet-lab validation, Bonbon achieved a 61.5% dose-response hit rate on de novo compounds. These compounds inhibited nine disease targets spanning seven protein families.
Key facts
- Bonbon was trained on 3.2 trillion tokens of paired protein-ligand sequences.
- The model shows emergent capabilities at functional, residue, and bond levels.
- It achieved a 61.5% dose-response hit rate in wet-lab validation.
- The model works at five orders of magnitude greater speed than co-folding methods.