sagawa/ReactionT5v1-forward

⚠️This is an old version of ReactionT5v2-forward. Prediction accuracy is worse.⚠️

Model Card for ReactionT5v1-forward

This is a ReactionT5 pre-trained to predict the products of reactions.

Model Sources

• Repository: https://github.com/sagawatatsuya/ReactionT5
• Paper: https://arxiv.org/abs/2311.06708

Uses

You can use this model for forward reaction prediction or fine-tune this model with your dataset.

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import AutoTokenizer, T5ForConditionalGeneration

tokenizer = AutoTokenizer.from_pretrained('sagawa/ReactionT5-product-prediction')
inp = tokenizer('REACTANT:COC(=O)C1=CCCN(C)C1.O.[Al+3].[H-].[Li+].[Na+].[OH-]REAGENT:C1CCOC1', return_tensors='pt')
model = T5ForConditionalGeneration.from_pretrained('sagawa/ReactionT5-product-prediction')
output = model.generate(**inp, min_length=6, max_length=109, num_beams=1, num_return_sequences=1, return_dict_in_generate=True, output_scores=True)
output = tokenizer.decode(output['sequences'][0], skip_special_tokens=True).replace(' ', '').rstrip('.')
output # 'O=S(=O)([O-])[O-].O=S(=O)([O-])[O-].O=S(=O)([O-])[O-].[Cr+3].[Cr+3]'

Training Details

Training Procedure

We used Open Reaction Database (ORD) dataset for model training. The command used for training is the following. For more information, please refer to the paper and GitHub repository.

python train.py \
    --epochs=100 \
    --batch_size=32 \
    --data_path='../data/all_ord_reaction_uniq_with_attr_v3.csv' \
    --use_reconstructed_data \
    --pretrained_model_name_or_path='sagawa/CompoundT5'

Results

Model	Training set	Test set	Top-1 [% acc.]	Top-2 [% acc.]	Top-3 [% acc.]	Top-5 [% acc.]
Sequence-to-sequence	USPTO	USPTO	80.3	84.7	86.2	87.5
WLDN	USPTO	USPTO	80.6 (85.6)	90.5	92.8	93.4
Molecular Transformer	USPTO	USPTO	88.8	92.6	–	94.4
T5Chem	USPTO	USPTO	90.4	94.2	–	96.4
CompoundT5	USPTO	USPTO	88.0	92.4	93.9	95.0
ReactionT5	-	USPTO	0.0 <85.0>	0.0 <90.6>	0.0 <92.3>	0.0 <93.8>

Performance comparison of Compound T5, ReactionT5, and other models in product prediction. The values enclosed in ‘<>’ in the table represent the scores of the model that was fine-tuned on 200 reactions from the USPTO dataset. The score enclosed in ‘()’ is the one reported in the original paper.

Citation

arxiv link: https://arxiv.org/abs/2311.06708

@misc{sagawa2023reactiont5,  
      title={ReactionT5: a large-scale pre-trained model towards application of limited reaction data}, 
      author={Tatsuya Sagawa and Ryosuke Kojima},  
      year={2023},  
      eprint={2311.06708},  
      archivePrefix={arXiv},  
      primaryClass={physics.chem-ph}  
}