Bonbon: 1.6B protein-ligand model hits 61.5% in wet-lab tests
TL;DR
- A 1.6B-parameter model named Bonbon, trained on ~3.2T paired protein-ligand sequence tokens with no structural coordinates, reports a 61.5% wet-lab dose-response hit rate.
- The wet-lab validation spans nine disease targets across seven protein families, with molecules the paper says had not previously been reported to inhibit those proteins.
- Three zero-shot capabilities — functional, residue-level, bond-level — appear only at the 1.6B scale even as pretraining loss falls smoothly across six model sizes.
A 1.6-billion-parameter model trained only on paired protein-ligand sequences hit a 61.5% dose-response rate in prospective wet-lab testing across nine disease targets, according to a single-author arXiv paper by Adi Mashiach.
The model is called Bonbon. It is described in the abstract as "a foundation model for molecular interactions trained on approximately 3.2 trillion tokens of paired protein-ligand sequences." From that one representation space, the paper claims three zero-shot capabilities at "hierarchical levels of resolution": functional mechanism of action, residue-level binding site localization "including orthosteric/allosteric discrimination," and bond-level covalent versus non-covalent engagement. All three, the paper says, "generalize to out-of-distribution complexes."
The scaling story is the unusual part. Mashiach ran "controlled scaling experiments across six model scales from 42M to 1.6B parameters" and reports that "every capability evaluated shows a sharp jump at the largest scale on continuous metrics, even as pretraining loss decreases smoothly." The paper reads this as evidence that "loss-based scaling laws alone are insufficient to predict when zero-shot emergent capabilities become available."
Training ran at a 1:2,000 parameter-to-token ratio, which the paper attributes to "the sparsity of paired biological interaction data."
The wet-lab number: a "61.5% dose-response hit rate on de novo compounds" across "nine disease targets spanning seven protein families," yielding molecules "not previously reported to inhibit these proteins, including for out-of-distribution targets." On binding and affinity prediction, the same frozen representations are said to run at "five orders of magnitude greater speed than co-folding methods."
The abstract does not name the targets, the comparator co-folding systems, or how a dose-response hit is scored.
Originally reported by paper
Read the original article →Original headline: NeurIPS 2026 Oral: Single-Author 1.6B Model on 3.2T Protein-Ligand Tokens Grows Zero-Shot Bio Capabilities, 61.5% Wet-Lab Hit Rate