A graph neural network based on torch geometric for aqueous solubility prediction
- Predict aqueous solubility of organic molecules using graph neural network, i.e., node (atom), edge (bond), and stacking layers, instead of chem descriptors
- Descriptors have been presented in the repo aqueous solubility prediction. Would like to compare and draw conclusion
- Challenges:
- multiclass classificaiton
- message passing includes not only node features but also edge's
- use wandb for hyperparameters sweeping
Dependencies
- PyTorch
- PyTorch-Geometric
- RDKit
- DeepChem
- Scikit-Learn
- Seaborn
- tqdm
Data Source
The well known AqSolDB: https://www.nature.com/articles/s41597-019-0151-1