from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("sagawa/ReactionT5v2-forward", return_tensors="pt")
model = AutoModelForSeq2SeqLM.from_pretrained("sagawa/ReactionT5v2-forward")

inp = tokenizer('REACTANT:COC(=O)C1=CCCN(C)C1.O.[Al+3].[H-].[Li+].[Na+].[OH-]REAGENT:C1CCOC1', return_tensors='pt')
output = model.generate(**inp, 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 # 'CN1CCC=C(CO)C1'

Training Details

Training Procedure

We used the Open Reaction Database (ORD) dataset for model training. In addition, we used USPTO_MIT dataset's test split to prevent data leakage. The command used for training is the following. For more information about data preprocessing and training, please refer to the paper and GitHub repository.

python

cd task_forward
python train.py \
    --output_dir='t5' \
    --epochs=100 \
    --lr=1e-3 \
    --batch_size=32 \
    --input_max_len=150 \
    --target_max_len=100 \
    --weight_decay=0.01 \
    --evaluation_strategy='epoch' \
    --save_strategy='epoch' \
    --logging_strategy='epoch' \
    --train_data_path='../data/preprocessed_ord_train.csv' \
    --valid_data_path='../data/preprocessed_ord_valid.csv' \
    --test_data_path='../data/preprocessed_ord_test.csv' \
    --USPTO_test_data_path='../data/USPTO_MIT/MIT_separated/test.csv' \
    --disable_tqdm \
    --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_MIT	USPTO_MIT	80.3	84.7	86.2	87.5
WLDN	USPTO_MIT	USPTO_MIT	80.6 (85.6)	90.5	92.8	93.4
Molecular Transformer	USPTO_MIT	USPTO_MIT	88.8	92.6	–	94.4
T5Chem	USPTO_MIT	USPTO_MIT	90.4	94.2	–	96.4
CompoundT5	USPTO_MIT	USPTO_MIT	86.6	89.5	90.4	91.2
ReactionT5 (This model)	-	USPTO_MIT	92.8	95.6	96.4	97.1
ReactionT5	USPTO_MIT	USPTO_MIT	97.5	98.6	98.8	99.0

Performance comparison of Compound T5, ReactionT5, and other models in product prediction.

Citation

@article{Sagawa2025,
  title   = {ReactionT5: a pre-trained transformer model for accurate chemical reaction prediction with limited data},
  author  = {Sagawa, Tatsuya and Kojima, Ryosuke},
  journal = {Journal of Cheminformatics},
  year    = {2025},
  volume  = {17},
  number  = {1},
  pages   = {126},
  doi     = {10.1186/s13321-025-01075-4},
  url     = {https://doi.org/10.1186/s13321-025-01075-4}
}