Emergent Biological Capabilities in a Foundation Model for Molecular Interactions
Bonbon achieved a reported 61.5% dose-response hit rate for newly designed compounds in wet-lab tests across nine disease targets.
The model trained on approximately 3.2 trillion tokens of paired protein–ligand sequences, without structural coordinates or task-specific supervision. It inferred mechanisms of action, located binding sites and distinguished covalent from non-covalent interactions, including on complexes outside its training distribution. Across six model sizes, all evaluated capabilities jumped sharply at 1.6 billion parameters despite smoothly improving training loss. The wet-lab tests yielded structurally novel molecules not previously reported to inhibit the tested proteins.
HF Daily Papers' note
The model trained on approximately 3.2 trillion tokens of paired protein–ligand sequences, without structural coordinates or task-specific supervision. It inferred mechanisms of action, located binding sites and distinguished covalent from non-covalent interactions, including on complexes outside its training distribution. Across six model sizes, all evaluated capabilities jumped sharply at 1.6 billion parameters despite smoothly improving training loss. The wet-lab tests yielded structurally novel molecules not previously reported to inhibit the tested proteins.
HF Daily Papers' note
score 5