https://github.com/YunjaeChoi/ggnnmols
Gated Graph Neural Network for Molecules
https://github.com/YunjaeChoi/ggnnmols
Last synced: about 2 months ago
JSON representation
Gated Graph Neural Network for Molecules
- Host: GitHub
- URL: https://github.com/YunjaeChoi/ggnnmols
- Owner: YunjaeChoi
- License: mit
- Created: 2019-04-10T11:14:20.000Z (about 6 years ago)
- Default Branch: master
- Last Pushed: 2019-04-10T11:41:47.000Z (about 6 years ago)
- Last Synced: 2025-03-16T11:11:26.753Z (2 months ago)
- Language: Jupyter Notebook
- Size: 21.4 MB
- Stars: 6
- Watchers: 2
- Forks: 3
- Open Issues: 0
-
Metadata Files:
- Readme: README.rst
- License: LICENSE
Awesome Lists containing this project
- awesome-graph-classification - [Python Reference
README
Gated Graph Neural Network for Molecules
***********************************************
Implementation of Gated graph neural network for molecules in tensorflow.
`Gated graph neural network paper `_
Dependencies
============
1. Rdkit
.. code:: shell
conda install -c rdkit rdkit
2. Tensorflow
- cpu-version
.. code:: shell
pip install tensorflow
- gpu-version
.. code:: shell
pip install tensorflow-gpu
3. others(for preprocessing): pandas, tqdm, sklearn
.. code:: shell
conda install pandas tqdm sklearn
Preprocessing
=============
1. Data
-------
`Tox21 Data `_
was used for training.
sdf data was converted to graph.(Only mols valid by rdkit was used.)
graph data consists:
- node features:
- atomic number(onehot)
- atomic number
- formal charge
- radical eletrons
- is aromatic(bool)
- hybridization(onehot)
- edge features:
- bond type(onehot)
- is in ring(bool)
2. preprocess.py
----------------
Does the following steps:
1. Extracts labels from csv file.
2. Check mols.
3. Get features.
4. Saves train/val/test data.
Training
========
1. Model
--------
Variant of Gated Graph Neural Network(GGNN)
- modifications:
- modified for undirected graph
- edge hidden features are used instead of weights per edge types
- node hidden features are updated via stacked GRUs instead of one GRU
2. train.ipynb
--------------
Does the following steps:
1. Loads preprcessed data
2. Trains with missing labels
3. Saves weights(best)
Result
======
1. saved_models
---------------
Current trained model(weights.best.hdf5) is in this folder.
2. evaluate.ipynb
-----------------
Model can be evaluated.
ex) Receiver Operating Characteristic curve of current saved model.
.. image:: https://github.com/YunjaeChoi/ggnnmols/blob/master/doc/image/roc.png