Molecular Property Prediction (Graph Property Prediction) on the ogbg-molhiv dataset. The task is to predict the target molecular properties as accurately as possible: wheter a molecule inhibits HIV replication or not.
ogb-molhiv is adapted from MoleculeNet and was introduced in Open Graph Benchmark: Datasets for Machine Learning on Graphs (NeurIPS 2020) and is part of the Open Graph Benchmark (OGB)
The nodes are atoms, and the edges of the graph are chemical bonds, exact feature desciptions can be found in the ogb code.
For more information (e.g. molecule name, chemical formula, figures) on the molecules in the dataset we provide a .ipynb notebook.
The model for this task is a Graph Isomorphism Network (GIN) from How powerful are Graph Neural Networks? which is a Graph Neural Network (GNN).
The model we use is the same as the one used as baseline on the official OGB Leaderboard by the authors of the dataset.
The performance is measured in ROC_AUC (receiver operating characteristic - area under curve): higher is better and the range of values is [0, 1].
Our model reaches a performance of 0.774 ± 0.0107. The search grid used to find the optimal hyperparameters can be found here.
In Semi-Supervised Classification with Graph Convolutional Networks, and on their Leaderboard the OGB team reports a Test ROC-AUC of 0.7558 ± 0.0140.
During development training was numerically unstable for learning rates > 1.0e-3 when using 16bit Automatic Mixed Precision (AMP). Make sure to try (a machine with) bf16 precision if you run into similar issues.