ogbg_molhiv

Graph Property Prediction for Open Graph Benchmark (OGB) Molhiv Dataset

The Open Graph Benchmark (OGB) is a collection of benchmark datasets, data loaders, and evaluators for graph machine learning. Here we complete the Graph Property Prediction task on molhiv dataset. Details can be found in our paper.

Results on ogbg-molhiv

Here, we demonstrate the following performance on the ogbg-molhiv dataset from Stanford Open Graph Benchmark (1.2.5)

Model	Test ROC-AUC	Validation ROC-AUC	Parameters	Hardware
Neural FingerPrints	0.8232 ± 0.0047	0.8331 ± 0.0054	2425102	Tesla V100 (32GB)

Reproducing results

Requirements

1. Create conda environment and install rdkit >= 2018.09.1 by:

   conda create -n ogbg_hiv python=3.6
   conda activate ogbg_hiv
   conda install -c conda-forge rdkit

2. Install GPU or CPU version of paddlepaddle == 1.8.4:

   pip install paddlepaddle==1.8.4             # cpu or
   pip install paddlpaddle-gpu == 1.8.4        # gpu

3. Install pgl == 1.2.1:

   pip install pgl==1.2.1

4. Install ogb for evaluation:

   pip install ogb

Training the model

To simply reproduce the results demonstrated above, run the following commands:

python extract_fingerprint.py --dataset_name ogbg-molhiv

CUDA_VISIBLE_DEVICES=0 python main.py --config hiv_config.yaml

The learned model parameters will be saved in ./outputs/task_name/, where the task_name is specified in hiv_config.yaml.

Then you can predict the learned morgan fingerprint vectors by running the following commands:

CUDA_VISIBLE_DEVICES=0 python main.py --config hiv_config.yaml --infer_model ./outputs/task_name/model_name

The predicted morgan fingerprint will be saved in ./dataset/ogbg-molhiv/soft_mgf_feat.npy.

To classify the property of moleculars by using the random forest classifier, run the following commands:

python random_forest.py --dataset_name ogbg-molhiv

Detailed hyperparameters

All the hyperparameters can be found in the hiv_config.yaml file.