Molecular foundation model Bonbon shows emergent zero-shot drug-design capabilities
OutputLabs · hf · 2026-11-16
Output Labs' Bonbon, a foundation model for molecular interactions trained on 3.2T tokens of paired protein-ligand sequences (NeurIPS 2026 Main Track Oral), exhibits emergent zero-shot capabilities.
Key findings
- Trained on protein-ligand sequences alone — no task supervision, no structural coordinates — it emerges three hierarchical zero-shot capabilities from one representation space: functional (mechanism of action), residue-level (binding site localization with orthosteric/allosteric discrimination), and bond-level (covalent vs non-covalent), all generalizing to out-of-distribution complexes.
- Controlled scaling experiments across six sizes (42M–1.6B) show pretraining loss decreasing smoothly while every capability jumps sharply at the largest scale — loss-based scaling laws alone can't predict when zero-shot emergence occurs.
- Frozen representations achieve SOTA binding and affinity prediction five orders of magnitude faster than co-folding methods.
- Prospective wet-lab validation across 9 disease targets (7 protein families) yields a 61.5% dose-response hit rate on de novo compounds, including structurally novel inhibitors for OOD targets.
More from Research
- METR researcher: frontier labs dodge hard science evals, searching under the streetlamp — rbhar90 · 2026-11-16
- Stanford's build-an-LLM-from-scratch course CS336 publishes full lectures, slides and code — stanfordnlp · 2026-11-16
- SpeedRunBench test: Inkling-2-Small clears SuperTux 1-1 in 1,961 frames — simonguozirui · 2026-11-16
- MEM-SQL: Memory-Empowered Text-to-SQL with HARE History-Aware Benchmark — hkuhk · 2026-11-16
- Grid Robotics Proposes Agentic Robotics Paradigm Built on Embodied System 2 — Grid Robotics · 2026-11-16
- LLM Peer Reviewers Fixate on Appendices, Unlike Humans, Study Finds — CincyNLP · 2026-11-16