How deepgraphlearning/gearnet works
GearNet is a PyTorch/TorchDrug protein-structure representation learning codebase. Users configure and launch pretraining or downstream experiments; datasets are constructed as protein graphs, encoded by geometric relational message passing, and trained/evaluated for protein-function or structure tasks. The graph captures documented pretraining objectives and benchmark families; configuration supports variants beyond the sampled implementation. Dataset and model details are sampled, while other TorchDrug-backed task and dataset wiring is documented or config-driven.
Main entry points: downstream.py Downstream runner · pretrain.py Pretrain runner · model.py GearNet encoder · pretrain Self-supervised objectives
Components 22
Experiment runs
- Pretrain runner
script/pretrain.py - Downstream runner
script/downstream.py - Experiment configs
config - Experiment setup
util.py
Protein data
- Protein datasets
gearnet/dataset.py - Fold3D loader
gearnet/dataset.py - Residue graphs
gearnet/dataset.py - Pretraining corpus
config/pretrain - Task benchmarks
config/downstream
Graph encoding
- GearNet encoder
gearnet/model.py - Sequence fusion
gearnet/model.py - Relational convolution
gearnet/layer.py - IEConv
gearnet/layer.py
Learning tasks
- Self-supervised objectives
config/pretrain - Multiview contrast
config/pretrain/mc_gearnet_edge.yaml - Residue prediction
config/pretrain/attr_gearnet_edge.yaml - Geometry prediction
config/pretrain - Task prediction
config/downstream - Model checkpoints
script/pretrain.py - Validation and test
script/downstream.py
Other components
- Researcher
- TorchDrug runtime
Connections 28
- Researcher → Pretrain runner: launches
- Researcher → Downstream runner: launches
- Pretrain runner → Experiment configs: loads config
- Downstream runner → Experiment configs: loads config
- Pretrain runner → Experiment setup: sets up
- Downstream runner → Experiment setup: sets up
- Pretrain runner → TorchDrug runtime: builds and trains
- Downstream runner → TorchDrug runtime: builds solver
- Experiment setup → TorchDrug runtime: constructs engine
- Pretrain runner → Pretraining corpus: loads dataset config
- Downstream runner → Task benchmarks: loads task config
- TorchDrug runtime → Protein datasets: constructs dataset
- Fold3D loader → Residue graphs: builds graphs
- Protein datasets → Residue graphs: provides graphs
- Protein datasets → TorchDrug runtime: splits samples
- TorchDrug runtime → GearNet encoder: invokes encoder
- GearNet encoder → Relational convolution: applies layers
- GearNet encoder → IEConv: optionally applies
- GearNet encoder → Sequence fusion: used by
- TorchDrug runtime → Self-supervised objectives: trains objective
- Self-supervised objectives → Multiview contrast: configures
- Self-supervised objectives → Residue prediction: configures
- Self-supervised objectives → Geometry prediction: configures
- Pretrain runner → Model checkpoints: saves weights
- Downstream runner → Task prediction: trains task
- Downstream runner → Validation and test: evaluates
- Task prediction → Task benchmarks: uses labels
- Downstream runner → Model checkpoints: loads weights