Foundation model predicts molecular interactions and drug effects fast
Emergent Biological Capabilities in a Foundation Model for Molecular Interactions
Machine Learning
Summary
Molecules like proteins and drugs interact in complex ways that are hard to predict. The authors created a large AI model called Bonbon that learned from huge amounts of sequence data without being told specific tasks. Bonbon can identify how drugs work, where they bind on proteins, and the type of chemical bonds involved, even for new molecules and proteins it has never seen before. It does this much faster than existing methods and was tested in real lab experiments, finding new drug candidates for several diseases.
What this means in practice
- •For drug discovery teams: Identify candidate drug molecules and their binding mechanisms quickly without experimental structures, speeding early drug screening for multiple disease targets.$Commercial implications: Enables rapid computational drug candidate identification and mechanism prediction, replacing slower structural methods, useful to pharmaceutical companies.
- •For computational chemists: Predict binding affinities and distinguish binding site types accurately using a single pretrained molecular interaction model for diverse proteins and ligands.
Authors
Adi Mashiach
Abstract
Foundation models trained on large unlabeled corpora develop emergent capabilities, tasks that go beyond the training objective. We extend this paradigm to multimodal biology with Bonbon, a foundation model for molecular interactions trained on approximately 3.2 trillion tokens of paired protein-ligand sequences. From a single representation space, Bonbon exhibits three zero-shot emergent capabilities at hierarchical levels of resolution: functional (mechanism of action), residue (binding site localization, including orthosteric/allosteric discrimination), and bond (covalent versus non-covalent engagement). All three capabilities generalize to out-of-distribution complexes. Bonbon was trained on protein-ligand sequences alone, without any task-specific supervision and without structural coordinates. We conducted controlled scaling experiments across six model scales from 42M to 1.6B parameters; every capability evaluated shows a sharp jump at the largest scale on continuous metrics, even as pretraining loss decreases smoothly. This demonstrates that loss-based scaling laws alone are insufficient to predict when zero-shot emergent capabilities become available. The capabilities reported here emerge after training at a 1:2,000 parameter-to-token ratio to accommodate the sparsity of paired biological interaction data. The same frozen representations achieve state-of-the-art binding and affinity prediction at five orders of magnitude greater speed than co-folding methods. In prospective wet-lab validation across nine disease targets spanning seven protein families, Bonbon achieves a 61.5% dose-response hit rate on de novo compounds, yielding structurally novel molecules not previously reported to inhibit these proteins, including for out-of-distribution targets.