How DeepMind's AlphaFold2 Predicts Protein Shapes

3-Second Blueprint (English)

1. AlphaFold2 aligns homologous sequences across species into a Multiple Sequence Alignment alongside a 2D amino-acid pair matrix.
2. The Evoformer transformer applies self-attention to iteratively exchange evolutionary constraints and spatial geometry updates.
3. A Structure transformer decodes 3D atomic coordinates and recycles predictions through the network with per-residue confidence.