AlphaFold 3
Diffusion-based biomolecular structure generator whose samples are ranked by a confidence formula with clash, disorder and chirality terms.
The loop
- Proposes
A Pairformer trunk and 3D atomic diffusion module generate joint all-atom coordinates for proteins, nucleic acids, ligands, and ions.
- Judged by
Samples are ranked by a formula of the model pTM and ipTM with clash, disorder and chirality terms; nothing is filtered out.
- Steers
Confidence ranking picks the top sample across diffusion rollouts; cross-distillation is a training-data step.
Headline result
Achieved 76.4% PoseBusters accuracy on protein-ligand complexes, far surpassing classical physics-based docking tools.
What this does not prove
Can hallucinate spurious secondary structure in disordered regions or produce occasional stereochemical clashes when diffusion samples conflict.
How it generalizes the harness
An all-atom SE(3) diffusion generator produces biomolecular complexes that are ranked by a confidence formula with clash, disorder and chirality terms.
Classification
| What it proposes | Protein structure |
|---|---|
| Generative model | Pairformer + 3D all-atom diffusion network |
| Algorithmic authority | Ranking formula of pTM and ipTM with clash, disorder and chirality terms |
| Feedback | Score |
| What gets updated | Population of candidates |
| When it runs | At inference time |
| Validity | Approximate physics (simulation, not lab results) |
| Evidence | Independently replicated |
| Lineage | Original system |