NPComposer
NPComposer is a conditional molecular generation model trained by fine-tuning GP-MoLFormer ibm-research/GP-MoLFormer-Uniq — a 46.8M parameter transformer decoder foundation model — on the COCONUT database containing over 700,000 experimentally validated natural products: https://github.com/shawnralyn/NPComposer
By providing class labels and molecular property information as special tokens during model fine-tuning, NPComposer allows for conditional natural product generation based on: NP biosynthesis pathway (7 pathways including Alkaloids, Terpenoids, Shikimates and Phenylpropanoids, etc.), NP superclass (70+ superclasses), presence or absence of glycoside, number of aromatic rings (0–22), QED drug-likeness (0–1), and synthetic accessibility score (1–10).
Model description
Intended uses & limitations
More information needed
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 5e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.03
- num_epochs: 8.0
Training results
| Training Loss | Epoch | Step | Validation Loss |
|---|---|---|---|
| 0.4069 | 1.0 | 27763 | 0.3820 |
| 0.3507 | 2.0 | 55526 | 0.3371 |
| 0.3247 | 3.0 | 83289 | 0.3142 |
| 0.3074 | 4.0 | 111052 | 0.2995 |
| 0.291 | 5.0 | 138815 | 0.2888 |
| 0.2765 | 6.0 | 166578 | 0.2803 |
| 0.2677 | 7.0 | 194341 | 0.2750 |
| 0.2569 | 8.0 | 222104 | 0.2718 |
Framework versions
- Transformers 4.44.2
- Pytorch 2.10.0+cu128
- Datasets 2.21.0
- Tokenizers 0.19.1
Hardware
1 x NVIDIA A40 48GB GPU