Papers
arxiv:2611.14504

Emergent Biological Capabilities in a Foundation Model for Molecular Interactions

Published on Nov 10
· Submitted by
Adi Mashiach
on Nov 16
Authors:

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.

Community

Paper author Paper submitter

Bonbon is a vertical foundation model for biology that independently learned the hidden relationships governing molecular interactions, just as LLMs learned math without being explicitly trained on it. Accepted to the NeurIPS 2026 main track as an Oral (top 0.4% of submissions).

Trained on sequences alone, with no structure, physics, task-specific labels, or supervision, Bonbon developed emergent biological capabilities as we scaled it from 42M to 1.6B parameters. Pretraining loss fell smoothly, while new capabilities appeared sharply at the largest scale:

  • What it does: predicts mechanism of action across six classes, including which molecules switch the same target on and which switch it off. Today this requires functional lab assays.
  • Where it happens: localizes the binding site and distinguishes orthosteric from allosteric engagement on the same target.
  • How it bonds: recognizes covalent binding and pinpoints the residue involved, without being told covalency is present.
  • Which atoms matter: identifies the active moiety at single-bond resolution, work that usually takes many rounds of making and testing analogs.

All capabilities were evaluated zero-shot on held-out proteins and molecules.

The same representation space reaches state-of-the-art performance on widely used binding and affinity benchmarks, and runs at 2,500 molecular interactions per second on a single H100, about 50,000× faster than a leading co-folding model.

In blinded dose-response assays at an external CRO, de novo molecules ranked by Bonbon achieved a 61.5% hit rate across nine disease targets in seven protein families, compared with the ~0.1% typical of high-throughput screening. Every molecule was less than 60% similar to anything in training. Some examples:

  • HDAC6: most known inhibitors rely on hydroxamate chemistry linked to toxicity. Bonbon found novel non-hydroxamate inhibitors.
  • MNK2: novel inhibitors outside the documented chemical space for this cancer target.
  • VEGFR2 and ELANE: both absent from training, yet Bonbon found confirmed inhibitors for each.

image4-targets

Why it matters

Most AI in drug discovery answers one narrow question: does a molecule fit a protein? But a molecule can fit perfectly and still do nothing, or the opposite of what you need. What it does, where it acts, how it holds, and which atoms drive it are usually answered through a long chain of separate assays and lab iterations.

Bonbon answers these questions together, in a single inference pass, at the scale biology demands. That collapses months of sequential experiments into one step, opens chemistry beyond known scaffolds, and puts the hardest drug programs within reach, including targets with no good starting points.

More broadly, it shows that the rules of molecular biology are learnable from sequence alone and emerge with scale. That is the foundation for moving from single interactions to pathways, cells, and understanding biology as a complex system.

Sign up or log in to comment

Get this paper in your agent:

hf papers read 2611.14504
Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash

Models citing this paper 0

No model linking this paper

Cite arxiv.org/abs/2611.14504 in a model README.md to link it from this page.

Datasets citing this paper 0

No dataset linking this paper

Cite arxiv.org/abs/2611.14504 in a dataset README.md to link it from this page.

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2611.14504 in a Space README.md to link it from this page.

Collections including this paper 0

No Collection including this paper

Add this paper to a collection to link it from this page.