Protein binding site classification with deep graph neural networks.
Graphsite is a software written in Python that reads a pocket (in .mol2) and compute its graph representation. In each generated graph, each atom represents a node. If the distance between two atoms are less than or equal to a threshold (default is 4.5 Angstrom), an undirected edge is formed between these two atoms. The edge attribute is the number of chemical bond(s) on this edge.
pip install graphsite
Graphsite is invented for applications which are based on graph neural networks (GNNs). However, the scope of this tool is not limited to deep graph learning. It can be used in any application that requires graph representaions of proteins/binding sites.
The main module of Graphsite is a callable class which reads 3 files as input:
from graphsite import Graphsitegs = Graphsite()node_feature, edge_index, edge_attr = gs(# path to the .mol2 file of pocketmol_path=mol_path,# path to the .profile file of pocket which# contains the sequence entropy node featureprofile_path=profile_path,# path to the .popsa file of pocket which contains# the solvent-accessible surface area node featurepop_path=pop_path)
The node_feature, edge_index, and edge_attr are numpy arrays. For more information about input paramters and output formats, see help:
help(Graphsite)
Below is an example where the output matrices of graphsite are used to create graphs for Pytorch-geometric.
from torch_geometric.data import Datadata = Data(x=node_feature,edge_index=edge_index,edge_attr=edge_attr))
For a complete deep learning example, please see Graphsite-classifier, where we build a graph classifier over 14 classes of binding pockets.
Graphsite is currently under peer review. Please cite this preprint if you find this package useful in your project :)
If you have any questions or suggestions, please submit an issue or pull request. Anyone is welcome to contribute :)