MLP (aka 'NeuralSym' by Segler and Waller, 2017)
We provide code and checkpoints for MLPs, which already yield decent performance. We trained two models MLP and MLP++ (using a larger fingerprint) over three datasets USPTO-50k, USPTO-rt, and USPTO-rd, using both the training and validation set. The exact model configurations are given in examples/configs.
Please note, the results vary slightly from the ones in our paper since:
- We updated the configurations.
- We used the cleaned USPTO-50k data from rxn-ebm.
- We used all templates also for the experiments over USPTO-rt and rd, for simplicity. In the paper, we dropped the ones occuring only once with MLP.
The performance for these models is as follows.
Trained Over USPTO-50k
USPTO-50k
| Model | mrr | top-1 | top-5 | top-10 | maxfrag-1 | maxfrag-5 | maxfrag-10 |
|---|---|---|---|---|---|---|---|
| MLP | 59.0 | 46.4 | 76.3 | 82.9 | 50.0 | 80.3 | 86.3 |
| MLP++ | 60.5 | 48.3 | 77.7 | 83.6 | 52.4 | 81.8 | 86.8 |
rt-1k
| Model | mrr | top-1 | top-5 | top-10 | maxfrag-1 | maxfrag-5 | maxfrag-10 | mss |
|---|---|---|---|---|---|---|---|---|
| MLP | 41.1 | 31.4 | 55.0 | 60.3 | 35.5 | 61.7 | 67.3 | 43.0 |
| MLP++ | 42.4 | 32.5 | 56.2 | 61.5 | 36.9 | 63.0 | 68.3 | 44.2 |
| rd-1k |
| Model | mrr | top-1 | top-5 | top-10 | maxfrag-1 | maxfrag-5 | maxfrag-10 | mss |
|---|---|---|---|---|---|---|---|---|
| MLP | 37.5 | 28.8 | 49.6 | 54.6 | 33.4 | 56.6 | 61.7 | 29.7 |
| MLP++ | 38.7 | 30.0 | 51.0 | 55.9 | 34.4 | 57.8 | 62.8 | 31.3 |
| Trained Over USPTO-rt |
rt-1k
| Model | mrr | top-1 | top-5 | top-10 | maxfrag-1 | maxfrag-5 | maxfrag-10 | mss |
|---|---|---|---|---|---|---|---|---|
| MLP | 48.9 | 37.4 | 65.2 | 72.6 | 41.6 | 70.5 | 77.6 | 55.6 |
| MLP++ | 49.3 | 37.9 | 65.6 | 72.1 | 42.2 | 70.8 | 77.1 | 54.8 |
| rt |
| Model | mrr | top-1 | top-5 | top-10 | maxfrag-1 | maxfrag-5 | maxfrag-10 | mss |
|---|---|---|---|---|---|---|---|---|
| MLP | 48.8 | 37.1 | 65.1 | 72.5 | 41.2 | 70.5 | 77.5 | 55.1 |
| MLP++ | 49.8 | 38.2 | 65.7 | 72.7 | 42.4 | 71.2 | 77.9 | 55.6 |
Trained Over USPTO-rd
rd-1k
| Model | mrr | top-1 | top-5 | top-10 | maxfrag-1 | maxfrag-5 | maxfrag-10 | mss |
|---|---|---|---|---|---|---|---|---|
| MLP | 51.5 | 40.3 | 66.8 | 73.5 | 45.0 | 71.5 | 78.2 | 48.7 |
| MLP++ | 52.3 | 41.4 | 67.5 | 73.5 | 45.9 | 72.9 | 78.7 | 47.8 |
| rd |
| Model | mrr | top-1 | top-5 | top-10 | maxfrag-1 | maxfrag-5 | maxfrag-10 | mss |
|---|---|---|---|---|---|---|---|---|
| MLP | 52.0 | 40.8 | 67.5 | 74.2 | 45.2 | 72.7 | 79.1 | 49.0 |
| MLP++ | 52.7 | 41.4 | 68.2 | 74.4 | 45.9 | 73.7 | 79.5 | 49.1 |