Fast Protein Structure Comparison through Effective Representation Learning with Contrastive Graph Neural Networks
Zenodo:
GraSR is implemented with Python3, so a Python3 (>=3.6) interpreter is required.
At first, download the source code of GraSR from GitHub:
$ git clone https://github.com/chunqiux/GraSR.git
Then, we recommend you to use a virtual environment, such as virtualenv, to install the dependencies of GraSR. If virtualenv is not available in your OS, try to install it as following:
$ pip3 install virtualenv
Then, create and activate the virtual environment as following:
# Higher python version is permitted
$ virtualenv gsr_env -p python3.6
$ source gsr_env/bin/activate
Then, install the dependencies as following:
$ pip install -r requirements.txt
When you want to quit the virtual environment, just:
$ deactivate
If you only want to generate descriptors for query structures, you can run the command as following:
$ python main.py -m saved_models -q example -o result
where '-m saved_models' means the directory path of models is './saved_models', '-q example' means the directory path of query structures is './example' and '-o result' means the output directory path is './result'.
The format of descriptor file is .pkl. you can check it in python3 as following:
$ python3
>>> import pickle
>>> with open("result/query_descriptors.pkl", "rb") as qd_file:
... d = pickle.load(qd_file)
... print(d)
'd' is a dictionary. Its keys are filenames of query structures and corresponding values are descriptors (Numpy.ndarray).
If you want to retrieve structural neighbors of the query structures, you can run the command as following:
$ python main.py -r -m saved_models -q example -k descriptors/scope_207_id40.pkl -o result
where '-r' means retrieval mode and '-k descriptors/scope_207_id40.pkl' means the file path of database is './descriptors/scope_207_id40.pkl'. An example of retrieval result is shown below:
Top-10 structural neighbors
sid Length-scaling cosine distance
d1ca4a1.ent 0.33015
d1lb6a_.ent 0.33941
d3ivva1.ent 0.36020
d2edma1.ent 0.38801
d4ca1a_.ent 0.38843
d2ed6a1.ent 0.42102
d2g1da1.ent 0.43577
d1qhva_.ent 0.43767
d5gv0a1.ent 0.44182
d4akma_.ent 0.44263
where 'sid' denotes the sid of the structural neighbors in the SCOPe and 'Length-scaling cosine distance' denotes the distance between the query structure and structural neighbors.
Our project is under Apache License. The moco.py is modified from MoCo/builder.py, which is under CC-BY-NC 4.0 license. The details can be referred from MoCo_LICENSE.
We also provide online retrieval service here. Our website follows a 'filter and refine' paradigm, which means it can provide more accurate result.