To benchmark the GTNN models introduced in our Topotein paper, we developed TopoteinWorkshop. This is a topological deep learning extension to the ProteinWorkshop framework for protein structure representation learning. This library substantially extends the original benchmark library to incorporate topological features and models for protein structure representation learning.

TopoteinWorkshop builds upon ProteinWorkshop to provide:

• Protein Combinatorial Complex data structure
• Novel Geometric Topological Neural Network architectures for protein structure learning, including TCPNet, GNN-TNN, and ETNN
• Benchmarking GTNN models, and evaluating them against existing GGNN models

# Create a conda environment using the provided environment file
uv sync

We use Python 3.10.13 and primarily run on Cambridge CSD3 Ampere GPU cluster. In most cases, you will need a GPU with high memory like A100 80G for training the larger models.

The main dependencies are managed through the conda environment file. Key dependencies include:

• PyTorch
• PyTorch Geometric
• PyTorch Lightning
• ProteinWorkshop
• Hydra

For dataset downloads, please refer to the ProteinWorkshop documentation. Topotein uses the same dataset formats and processing pipelines as ProteinWorkshop.

We use Hydra to handle configurations. Check ProteinWorkshop/proteinworkshop/config for full details of possible configurations.

Example command to run our model:

python ./ProteinWorkshop/proteinworkshop/train.py \
  encoder=tcpnet_v0 \
  task=multiclass_graph_classification \
  dataset=fold_superfamily \
  features=ca_bb_sse_3di \
  +aux_task=nn_sequence_3di \
  dataset.datamodule.dataset_fraction=1 \
  logger=wandb \
  trainer.max_epochs=150

Topotein implements several topological neural network architectures:

• TCPNet (tcpnet_v1 for using protein-level message passing, tcpnet_v0 for not using this channel)
• GVP-TNN (tvp)
• ETNN (etnn for our implementation optimized for protein combinatorial complex, etnn_original for the original implementation)

The main content of our framework are stored in the topotein folder within the ProteinWorkshop directory. This includes:

• GTNN models in topotein/models/
• Protein combinatorial complex and topological featurisation in topotein/features/

This project is licensed under the MIT License - see the LICENSE file for details.

• This project builds upon ProteinWorkshop
• We thank the Cambridge CSD3 for computational resources