MODEL LEARNING PATH
AlphaFold 3
AlphaFold 3 represents a molecular complex at both token and atom resolution. A single/pair trunk conditions an atom–token–atom denoiser that repeatedly refines noisy coordinates.
Loading the interactive page…
Introduced a unified diffusion-based structure model across several biomolecular types. Explicit reproductions such as Protenix demonstrate architectural influence, although its historical impact has had less time to develop than AlphaFold2's.
Chapters
Describe the complex
Prepared molecular, reference-conformer, template and MSA features describe the input entities.
Trace molecular identities and reference geometry into token and pair context.
Reason over tokens and pairs
The Pairformer develops the context used throughout denoising.
Combine a pair update with attention while preserving token and pair axes.
Denoise one coordinate sample
EDM scaling wraps the local atom encoder, conditioned token transformer and atom decoder.
Follow atom-to-token-to-atom denoising and its EDM scaling wrapper.
Sample and assess
Repeated denoiser calls advance the noise schedule; confidence calculations assess the resulting complexes.
Separate a sampling step from confidence and ranking calculations.
All exercises in this guide
- Token Embedding Lookup
- AlphaFold 3 Relative Position Buckets
- Affine Projection
- ReLU
- Atom Layout Gather
- Reference Atom Pair Features
- Masked Sum
- Masked Softmax
- Masked Mean
- Layer Normalization
- AlphaFold 3 Adaptive Normalization
- Atom–Token Broadcast and Pool
- AlphaFold 3 Atom Attention Encoder
- Stable vector norms, normalization, dot and cross features
- Batched rigid transforms
- Right-handed local frame from three points
- Masked pairwise distances and hard distance-bin targets
- AlphaFold 3 Template Pair Features
- Template Pair Mean
- Chain, entity and token-relative position buckets
- Categorical and numeric molecular feature encoding
- Feedforward Block
- Template geometry and pair features
- Multi-head attention with masks and additive bias
- Template-axis attention and mean aggregation
- MSA pair-weighted averaging
- Masked outer-product mean
- Triangle multiplicative update: incoming and outgoing
- Triangle attention: starting and ending node
- AlphaFold 3 Recycling Update
- Recycling representations and clean-template state
- Gaussian forward corruption and score targets
- EDM coordinate-denoiser preconditioning
- SiLU
- Gated Feedforward Block
- Fourier noise features and learned conditioning projection
- LayerNorm and RMSNorm from reductions
- Affine projection and embedding lookup
- Activations and gated feed-forward transitions
- Adaptive LayerNorm and RMS condition modulation
- Noisy Atom Pair Features
- Trunk-Conditioned Atom Pair Features
- Single-to-pair features and atom reference conditioning
- Ragged segment sum and mean
- Atom-layout gather and masked atom-to-token pooling
- Conditioned pair-bias attention and gated transition block
- AlphaFold 3 Atom Attention Decoder
- Rotary positions for queries and keys
- Local atom attention, encoder pooling and decoder skip/readout
- Shared rigid coordinate augmentation and Haar rotations
- Weighted proper rigid alignment
- Power-schedule EDM initialization and sampler step
- Normalized Linear pLDDT Head
- Normalized Resolved-Atom Head
- Distogram Linear Head
- PAE Linear Head
- PDE Linear Head
- pLDDT Confidence Summary
- Predicted TM Score
- Interface TM Score
- Directed Chain-Pair TM Scores
- Bidirectional Chain-Pair TM Scores
- Relative Accessible Surface Area
- TM Ranking Score
- Polymer Clash Summary
- AlphaFold 3 Ranking Score
- Learned structural confidence heads and model-aware ranking
- Peptide geometry, steric violations and polymer clashes
- Local distance difference test (lDDT)
- Rigid Permutation Alignment
- Symmetry assignment: rigid fits and ambiguous atom renaming
- Geometric confidence targets and observation masks