Graph Attention Site Prediction (GrASP)

Publication

https://pubs.acs.org/doi/10.1021/acs.jcim.3c01698

Colab

Fetch a PDB file and try GrASP on it in our Colab demo.

Download Datasets

In each dataset, ready_to_parse_mol2.zip contains the minimal structure files necessary to predict and evaluate binding sites with a general method, while processed.zip contains the PyTorch Geometric graphs used to run GrASP.

GrASP sc-PDB

GrASP COACH420

GrASP HOLO4K

How to Run

Currently, only production mode on a pre-trained model is supported until datasets are online.

• Build the conda environment by running

mamba create -n grasp python==3.7.10

mamba install conda-forge::cython
mamba install conda-forge::openbabel=2.4.1
mamba install conda-forge::rdkit
mamba install conda-forge::mdtraj
mamba install conda-forge::mdanalysis

pip install networkx==2.5 ```

* Move protein structures to `./benchmark_data_dir/production/unprocessed_inputs/`. Heteroatoms do not need to be removed, they will be cleaned during parsing.
* Load `ob` and parse the structures into graphs.

python3 parse_files.py production

* Run GrASP over the protein graphs.

python3 infer_test_set.py

* Paint structures with GrASP predictions in the b-factor column.

conda deactivate; conda activate ob python3 color_pdb.py

## Supported Formats
PDB and mol2 formats are supported and validated. Other formats supported by both MDAnalysis and OpenBabel 2.4.1 may be working but have not been tested.