How proteindesignlab/slae works
SLAE is a pretrained, all-atom protein-structure representation and reconstruction framework. Its principal workflow loads a structure into batched protein data, featurizes atoms and local graph edges, encodes the graph into atom/residue/graph embeddings, and optionally decodes residue embeddings into full-atom coordinates for PDB output. The graph also captures the dataset, tensor-conversion, and neural-network building-block layers. Some implementation details are unsampled; those relationships are shown only where the README or excerpts support them.
Main entry points: inference.py Inference API · datamodule.py Data module · dataset.py Structure dataset · dataloader.py Protein loader
Components 13
Structure input
- Structure dataset
SLAE/datasets/dataset.py - Data module
SLAE/datasets/datamodule.py - Protein loader
SLAE/datasets/dataloader.py
Graph representation
- Atom tensors
SLAE/io/atom_tensor.py - Atom representation
SLAE/features/fa_representation.py - Graph featurizer
SLAE/features/graph_featurizer.py
Autoencoder models
- Protein encoder
SLAE/model/encoder.py - Equivariant blocks
SLAE/nn - All-atom decoder
SLAE/model/decoder.py
Inference output
- Inference API
SLAE/inference.py - Reconstructed coordinates
- PDB writer
SLAE/io/write_pdb.py
Other components
- Researcher
Connections 14
- Researcher → Inference API: requests encode
README.md - Data module → Structure dataset: creates datasets
SLAE/datasets/datamodule.py - Data module → Protein loader: provides batches
SLAE/datasets/datamodule.py - Structure dataset → Atom tensors: converts atoms
SLAE/datasets/dataset.py - Protein loader → Graph featurizer: batches data for
README.md - Graph featurizer → Atom representation: transforms representation
SLAE/features/graph_featurizer.py - Atom representation → Atom tensors: extracts atoms
SLAE/features/fa_representation.py - Inference API → Protein encoder: runs encode
README.md - Protein encoder → Equivariant blocks: uses building blocks
SLAE/model/encoder.py - Inference API → All-atom decoder: runs decode
README.md - All-atom decoder → Reconstructed coordinates: reconstructs coordinates
README.md - Reconstructed coordinates → PDB writer: writes structure
README.md - Researcher → Data module: selects structure data
README.md - Inference API → Reconstructed coordinates: returns embeddings and coordinates
README.md