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.

PropertyDetail
TypeResearch 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


MOCs