AlphaFold 2

Its blind CASP14 result in 2020 established a major advance in protein-structure prediction; the full method appeared in Nature in 2021. The 2024 Chemistry Nobel recognition explicitly identifies AlphaFold2 as the breakthrough.

Chapters

1. Gather evidence

   Sequence, aligned homologs and templates initialize the MSA and pair representations.

   Build initial residue and pair features from sequence evidence.

   • Affine Projection
   • Chain-Aware Relative Position Buckets
   • MSA cluster profiles and deletion features

2. Exchange information

   The Evoformer alternates MSA reasoning with geometric reasoning over residue pairs.

   Explain how aligned sequences update pair features and how pairs guide attention.

   • Masked outer-product mean
   • MSA row attention with pair bias
   • Triangle multiplicative update: incoming and outgoing

3. Build and refine structure

   The structure module updates residue frames, reconstructs atoms and feeds previous predictions into another network pass.

   Apply a residue-frame update and identify what a later recycle pass receives.

   • Batched rigid transforms
   • Local Quaternion Frame Update
   • AlphaFold 2 Recycling Update

4. Read outputs and learn

   Confidence heads accompany the predicted structure. FAPE compares it with a target only on the training branch.

   Distinguish predicted confidence from a training loss measured in local frames.

   • pLDDT Confidence Summary
   • Express Points in Local Frames
   • Reduce Local Frame Point Errors

All exercises in this guide

• Token Embedding Lookup
• Chain-Aware Relative Position Buckets
• Affine Projection
• Stable vector norms, normalization, dot and cross features
• Batched rigid transforms
• Right-handed local frame from three points
• Masked Sum
• Masked Mean
• Masked pairwise distances and hard distance-bin targets
• AlphaFold 2 Template Pair Features
• Masked Softmax
• Template Pair Cross-Attention
• Chain, entity and token-relative position buckets
• Affine projection and embedding lookup
• Ragged segment sum and mean
• MSA cluster profiles and deletion features
• Template geometry and pair features
• Multi-head attention with masks and additive bias
• Template-axis attention and mean aggregation
• Layer Normalization
• Global-column MSA attention with shared keys and values
• MSA row attention with pair bias
• MSA column attention
• Masked outer-product mean
• Feedforward Block
• Triangle multiplicative update: incoming and outgoing
• Triangle attention: starting and ending node
• Invariant point attention (IPA)
• Local Quaternion Frame Update
• Normalized-quaternion backbone updates
• Learned torsions and hierarchical atom reconstruction
• AlphaFold 2 Recycling Update
• Recycling representations and clean-template state
• ReLU
• ReLU pLDDT Head
• Resolved-Atom Head
• Distogram Linear Head
• PAE Linear Head
• pLDDT Confidence Summary
• Predicted TM Score
• LayerNorm and RMSNorm from reductions
• Local distance difference test (lDDT)
• Atom Layout Gather
• Atom-layout gather and masked atom-to-token pooling
• Interface TM Score
• TM Ranking Score
• Learned structural confidence heads and model-aware ranking
• Express Points in Local Frames
• Reduce Local Frame Point Errors
• Ambiguous Atom Renaming
• Frame-aligned point error (FAPE)
• Weighted proper rigid alignment
• Rigid Permutation Alignment
• Symmetry assignment: rigid fits and ambiguous atom renaming
• Geometric confidence targets and observation masks
• Peptide Geometry Violation
• Within-Residue Distance Violation
• Between-Residue Clash Loss
• Peptide geometry, steric violations and polymer clashes