used parts of the code from: https://github.com/MolecularAI/Chemformer
usage: specify parameters, datasets and models used in graph/config.yaml and run python graph/main.py with
flags --vocab_path to specify the vocabulary for the NLP part and --model_path to specify path to the
pretrained chemformer model (https://az.app.box.com/s/7eci3nd9vy0xplqniitpk02rbg9q2zcq)
Results on freesolv:
| Regression Model | Fusion | Dataset | RMSE Mean (test) | RMSE Std (test) | RMSE (test) | R2 (test) | Flags |
|---|---|---|---|---|---|---|---|
| Transformer + GCN wo Reconstruction | Concat -> 2 dense layers with ReLU | freesolv | 1.450 | 0.221 | 1.450 +- 0.221 | 0.901 +- 0.030 | run.n_runs=10 |
| Transformer + GCN w/ Reconstruction | Concat -> 2 dense layers with ReLU | freesolv | 1.451 | 0.103 | 1.451 +- 0.103 | 0.903 +- 0.014 | run.n_runs=10 |
| GCN wo Reconstruction | freesolv | 1.803 | 0.060 | 1.803 +- 0.060 | 0.849 +- 0.010 | run.n_runs=10 | |
| GCN w/ Reconstruction | freesolv | 1.904 | 0.044 | 1.904 +- 0.044 | 0.833 +- 0.008 | run.n_runs=10 | |
| Transformer wo Reconstruction | freesolv | 2.071 | 0.291 | 2.071 +- 0.291 | 0.796 +- 0.057 | run.n_runs=10 | |
| Transformer w/ Reconstruction | freesolv | 2.104 | 0.381 | 2.104 +- 0.381 | 0.787 +- 0.084 | run.n_runs=10 |