Bonbon shows zero-shot molecular skills without structural training
Original title: Emergent Biological Capabilities in a Foundation Model for Molecular Interactions
A 1.6B model trained only on protein-ligand sequences reports binding-site, mechanism, and bond predictions plus a 61.5% wet-lab hit rate.
Researchers describe Bonbon, a foundation model trained on about 3.2 trillion tokens of paired protein-ligand sequences and no structural coordinates or task labels. From one representation it reports zero-shot functional, residue, and bond-level predictions that generalize to out-of-distribution complexes. Controlled scaling from 42M to 1.6B parameters shows capability jumps only at the largest scale. Frozen representations are said to match state-of-the-art binding and affinity prediction much faster than co-folding, with a 61.5% dose-response hit rate across nine disease targets.
Unconfirmed: Capabilities, speed, and the 61.5% hit rate are author-reported and not independently confirmed here.
- Training used a 1:2,000 parameter-to-token ratio for sparse interaction data.
- Every evaluated capability jumped at 1.6B parameters while pretraining loss fell smoothly.
- Authors say inference is five orders of magnitude faster than co-folding methods.
- Prospective tests covered seven protein families and structurally novel inhibitors.
From the original article
1 min read at arxiv.orgFoundation 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.
By Adi Mashiach · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.