How google-deepmind/alphafold3 works
AlphaFold 3 is a biomolecular-structure prediction inference package. Its principal workflow accepts molecular input, optionally searches sequence and template databases, prepares model features, runs neural-network inference, and post-processes predictions. The README documents data-pipeline and inference as independently switchable stages. Edges are limited to README statements and references visible in the SOURCE INDEX; unsampled implementation details and output-format wiring are not asserted.
Main entry points: pipeline.py MSA pipeline · features.py Feature preparation · folding_input.py Fold input · templates.py Template search
Components 20
Input and chemistry
- Fold input
src/alphafold3/common/folding_input.py - Component chemistry
src/alphafold3/constants/chemical_components.py
Sequence and templates
- MSA pipeline
src/alphafold3/data/pipeline.py - Alignment tools
src/alphafold3/data/tools - Template search
src/alphafold3/data/templates.py - Alignment parsing
src/alphafold3/data/parsers.py
Feature and model inference
- Feature preparation
src/alphafold3/model/features.py - Atom layout
src/alphafold3/model/atom_layout/atom_layout.py - Inference pipeline
src/alphafold3/model/pipeline/pipeline.py - Neural model
src/alphafold3/model/model.py - Confidence scoring
src/alphafold3/model/confidences.py
Structure processing
- Prediction post-processing
src/alphafold3/model/post_processing.py - Structure parsing
src/alphafold3/structure/parsing.py - Structure operations
src/alphafold3/structure/structure.py - mmCIF support
src/alphafold3/structure/mmcif.py
Other components
- Prediction user
- Sequence databases
- Structure database
- Model parameters
src/alphafold3/model/params.py - Predicted structure
Connections 21
- Prediction user → Fold input: supplies input
README.md - Fold input → MSA pipeline: enters search
src/alphafold3/data/pipeline.py - MSA pipeline → Alignment tools: runs alignments
- MSA pipeline → Sequence databases: searches sequences
- MSA pipeline → Template search: gets templates
src/alphafold3/data/pipeline.py - Template search → Structure database: searches structures
- Template search → Alignment parsing: uses parsers
src/alphafold3/data/templates.py - MSA pipeline → Alignment parsing: parses alignments
- Fold input → Feature preparation: supplies molecular input
src/alphafold3/model/features.py - Feature preparation → Atom layout: uses atom layout
src/alphafold3/model/features.py - Feature preparation → Component chemistry: uses components
src/alphafold3/model/features.py - Feature preparation → MSA pipeline: uses MSA and templates
src/alphafold3/model/features.py - Feature preparation → Inference pipeline: provides features
src/alphafold3/model/pipeline/pipeline.py - Inference pipeline → Neural model: runs inference
- Neural model → Model parameters: loads parameters
- Neural model → Confidence scoring: produces confidence data
- Inference pipeline → Prediction post-processing: post-processes prediction
- Prediction post-processing → Predicted structure: returns prediction
- Structure parsing → mmCIF support: parses mmCIF
src/alphafold3/structure/parsing.py - Structure parsing → Structure operations: builds structures
src/alphafold3/structure/parsing.py - Structure operations → mmCIF support: uses mmCIF utilities
src/alphafold3/structure/structure.py