GNN prediction of molecular solubility
- Base model
- SchNet, PyTorch 2.3 model.py
- Method
- Original design, improving on the baseline in [1] ref-1.pdf
- Parameters
- Depth and learning rate from 5-fold CV cv.log
- Rationale
- Scaffold split to avoid data leakage split.py
- All values
- Test RMSE 0.58, per-molecule predictions pred.csv
- Convergence
- RMSE within 3% across 5 seeds loss.png