Iambackup/ReactionT5v2-retrosynthesis

1---2language:3- en4license: mit5tags:6- chemistry7- SMILES8- retrosynthesis9datasets:10- ORD11metrics:12- accuracy13---14 15# Model Card for ReactionT5v2-retrosynthesis16 17This is a ReactionT5 pre-trained to predict the reactants of reactions. You can use the demo [here](https://huggingface.co/spaces/sagawa/ReactionT5_task_retrosynthesis).18 19 20### Model Sources21 22<!-- Provide the basic links for the model. -->23 24- **Repository:** https://github.com/sagawatatsuya/ReactionT5v225- **Paper:** https://jcheminf.biomedcentral.com/articles/10.1186/s13321-025-01075-426- **Demo:** https://huggingface.co/spaces/sagawa/ReactionT527 28## Uses29 30<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->31You can use this model for retrosynthesis prediction or fine-tune this model with your dataset.32 33 34## How to Get Started with the Model35 36Use the code below to get started with the model.37 38```python39from transformers import AutoTokenizer, AutoModelForSeq2SeqLM40 41tokenizer = AutoTokenizer.from_pretrained("sagawa/ReactionT5v2-retrosynthesis", return_tensors="pt")42model = AutoModelForSeq2SeqLM.from_pretrained("sagawa/ReactionT5v2-retrosynthesis")43 44inp = tokenizer('CCN(CC)CCNC(=S)NC1CCCc2cc(C)cnc21', return_tensors='pt')45output = model.generate(**inp, num_beams=1, num_return_sequences=1, return_dict_in_generate=True, output_scores=True)46output = tokenizer.decode(output['sequences'][0], skip_special_tokens=True).replace(' ', '').rstrip('.')47output # 'CCN(CC)CCN=C=S.Cc1cnc2c(c1)CCCC2N'48```49 50## Training Details51 52### Training Procedure 53 54<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->55We used the [Open Reaction Database (ORD) dataset](https://drive.google.com/file/d/1JozA2OlByfZ-ILt5H5YrTjLJvSvD8xdL/view?usp=drive_link) for model training. In addition, we used [USPTO_50k dataset](https://yzhang.hpc.nyu.edu/T5Chem/index.html)'s test split to prevent data leakage.56The command used for training is the following. For more information about data preprocessing and training, please refer to the paper and GitHub repository.57 58```python59cd task_retrosynthesis60python train.py \61    --output_dir='t5' \62    --epochs=80 \63    --lr=2e-4 \64    --batch_size=32 \65    --input_max_len=100 \66    --target_max_len=150 \67    --weight_decay=0.01 \68    --evaluation_strategy='epoch' \69    --save_strategy='epoch' \70    --logging_strategy='epoch' \71    --train_data_path='../data/preprocessed_ord_train.csv' \72    --valid_data_path='../data/preprocessed_ord_valid.csv' \73    --test_data_path='../data/preprocessed_ord_test.csv' \74    --USPTO_test_data_path='../data/USPTO_50k/test.csv' \75    --pretrained_model_name_or_path='sagawa/CompoundT5'76```77 78### Results79 80| Model                | Training set              | Test set | Top-1 [% acc.] | Top-2 [% acc.] | Top-3 [% acc.] | Top-5 [% acc.] |81|----------------------|---------------------------|----------|----------------|----------------|----------------|----------------|82| Sequence-to-sequence | USPTO_50k                 | USPTO_50k    | 37.4           | -           | 52.4           | 57.0           |83| Molecular Transformer| USPTO_50k                 | USPTO_50k    | 43.5           | -           | 60.5              | -           |84| SCROP                | USPTO_50k                 | USPTO_50k    | 43.7           | -          | 60.0           | 65.2           |85| T5Chem               | USPTO_50k                 | USPTO_50k    | 46.5           | -           | 64.4              | 70.5           |86| CompoundT5           | USPTO_50k                 | USPTO_50k    | 44,2           | 55.2           | 61.4           | 67.3           |87| [ReactionT5 (This model)](https://huggingface.co/sagawa/ReactionT5v2-retrosynthesis) | -                       | USPTO_50k    | 13.8     | 18.6     | 21.4     | 26.2     |88| [ReactionT5](https://huggingface.co/sagawa/ReactionT5v2-retrosynthesis-USPTO_50k)           | USPTO_50k                       | USPTO_50k    | 71.2     | 81.4     | 84.9     | 88.2     |89 90Performance comparison of Compound T5, ReactionT5, and other models in product prediction.91 92## Citation93 94<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->95```96@article{Sagawa2025,97  title   = {ReactionT5: a pre-trained transformer model for accurate chemical reaction prediction with limited data},98  author  = {Sagawa, Tatsuya and Kojima, Ryosuke},99  journal = {Journal of Cheminformatics},100  year    = {2025},101  volume  = {17},102  number  = {1},103  pages   = {126},104  doi     = {10.1186/s13321-025-01075-4},105  url     = {https://doi.org/10.1186/s13321-025-01075-4}106}107```