A new foundation model called Bonbon can predict three different aspects of a drug’s interaction with a protein from sequence data alone. The model learns a single representation space that reveals the drug’s mechanism of action, where it binds – whether it sits in the main pocket or in an alternative allosteric site – and whether it forms a covalent bond. These predictions appear without any task‑specific training or structural coordinates.

Bonbon is a transformer trained on about 3.2 trillion paired protein‑ligand tokens, using self‑supervision. No extra labels or structures were needed during training.

Researchers tested the model at six sizes, ranging from 42 million to 1.6 billion parameters. Across all scales the training loss fell smoothly, but each of the three skills showed a sharp jump only when the largest 1.6 billion‑parameter model was used. This shows that the usual loss‑based scaling laws cannot predict when such zero‑shot abilities emerge.

The study’s promise is tempered by limits. The model currently handles only broad distinctions: it tells if a drug acts orthosterically or allosterically, and whether it reacts covalently, but it does not provide precise 3D coordinates or finer details. All reported results come from frozen representations; the model was not fine‑tuned for any downstream task, so the shown capabilities represent a minimum of its potential.