LN — AlphaFold2 (Jumper et al., 2021)
Full Reference: Jumper, J., Evans, R., Pritzel, A., et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873), 583–589. https://doi.org/10.1038/s41586-021-03819-2
Pass 1 — Bird’s Eye
Main Claim
AlphaFold2 predicts protein structures with near-experimental accuracy even without structural homologues, winning CASP14 by a large margin.
| Property | Detail |
|---|---|
| Type | Research paper — Computational Biology / AI |
| Relevance to PUMA | ⭐⭐ Medium — canonical example of AI achieving breakthrough results in science; context for PUMA’s positioning of LLM agents as equivalent transformers for PM tasks |
Pass 2 — Key Content
System Design
- Transformer-based architecture trained on PDB structures + evolutionary sequence data
- Key innovation: attention mechanism over Multiple Sequence Alignments (MSAs) enables structural inference
- Output: 3D atom coordinates with confidence scores (pLDDT)
Results
- CASP14: GDT score ~90 (near-experimental quality threshold ~90)
- Predicted 98.5% of human proteome (AlphaFold DB, 200M+ structures)
- Enabled: drug discovery, enzyme design, understanding of disease mechanisms
Knowledge Generated
- Not just prediction: AlphaFold inferences revealed folding pathway insights and structural principles not explicitly programmed
- Provides structural hypotheses that guide new experimental lines
Pass 3 — PUMA Relevance
PUMA Analogy
AlphaFold demonstrates the “AI as scientific model” level in LeCun’s framework. PUMA occupies a similar position for PM:
- AlphaFold: input (sequence) → output (structure) with implicit physical knowledge
- PUMA triage agent: input (issue text) → output (priority) with implicit PM domain knowledge
Both use learned representations to solve domain-specific prediction tasks that previously required human expertise.
APA7 Citation
Jumper, J., Evans, R., Pritzel, A., et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873), 583–589. https://doi.org/10.1038/s41586-021-03819-2
Related Notes
- PN-AI-Scientific-Knowledge-Generation — synthesised permanent note
- PN-PUMA-within-AgenticScience-Trajectory — PUMA as Level 2 analogue
- PN-Agentic-Science-Paradigm — Agentic Science context
- LN-Jain-2023-GNoME-Materials
- LN-Lam-2023-GraphCast — sister L2 example
- LN-Zhang-2025-AgenticScienceSurvey — survey context