Graph Auto-encoder [1] implemented with DGL by Shion Honda.
Official implementation by the authors is here (TensorFlow, Python 2.7).
Unlike other implementations, this repository supports inductive tasks using molecular graphs (ZINC-250k), showing the power of graph representation learning with GAE.
You need PyTorch and DGL at least and the rest to try inductive settings with molecular graphs.
PyTorch
DeepGraphLibrary
RDKit
dill
tqdm
Reproduce the results of the paper [1] by the following command.
$ python train_transductive.py --dataset cora
You can switch the dataset to use by assigning to the --dataset option one from cora/citeseer/pubmed.
This repository supports learning graph representations of molecules in the ZINC-250k dataset (or any unlabeled SMILES dataset). Run pre-training by the following commands.
$ python prepare_data.py # download and preprocess zinc dataset
$ python train_inductive.py --hidden_dims 32 16 # pre-train GAE
The ZINC-250k is a subset of ZINC dataset and can be obtained easily by, for example, Chainer Chemistry.
Interestingly, I found GAE also works in inductive settings even though it was not tested in the original paper [1].
Is learned feature through pre-training really useful for predicting molecular properties? Let's check with simple examples. Here I use ESOL (solubility regression) dataset from [2], which can be downloaded here.
| Feature + Model | RMSE | R2 |
|---|---|---|
| GAE + Ridge | 1.813 | 0.585 |
| GAE + MLP | 1.216 | 0.732 |
| GAE + Random Forest | 1.424 | 0.688 |
| ECFP + Ridge | 2.271 | 0.480 |
| ECFP + MLP | 2.034 | 0.549 |
| ECFP + Random Forest | 1.668 | 0.643 |
ECFP is a hash-based binary feature of molecules (
GAE feature is a concatenation of mean, sum, and max aggregation of the hidden vector
GAE performs better than ECFP in all the combination with three regression models: Ridge, Multi-layer perceptron, and Random Forest.
[1] Thomas N. Kipf and Max Welling. "Variational Graph Auto-Encoders." NIPS. 2016.
[2] Zhenqin Wu, Bharath Ramsundar, Evan N. Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S. Pappu, Karl Leswing, Vijay Pande. "MoleculeNet: A Benchmark for Molecular Machine Learning", Chemical Science. 2018.