SMILES-based Transformer Encoder-Decoder (SMI-TED)

This repository provides a HuggingFace-compatible version of the SMI-TED model, a SMILES-based Transformer Encoder-Decoder for chemical language modeling.

📦 Forked Resources

Forked GitHub: bisect-group/materials-smi-ted-fork
Forked HuggingFace: bisectgroup/materials-smi-ted-fork

🏷️ Original Resources

Original GitHub: IBM/materials (smi_ted)
Original HuggingFace: ibm/materials.smi-ted
Publication: A Large Encoder-Decoder Family of Foundation Models for Chemical Language

🚀 Usage

pip install smi-ted

import torch
import smi_ted
from transformers import AutoConfig, AutoModel, AutoTokenizer

# Load config, tokenizer, and model from HuggingFace Hub
config = AutoConfig.from_pretrained("bisectgroup/materials-smi-ted-fork")
tokenizer = AutoTokenizer.from_pretrained("bisectgroup/materials-smi-ted-fork")
model = AutoModel.from_pretrained("bisectgroup/materials-smi-ted-fork")

# Link tokenizer to model (required for SMILES reconstruction)
model.smi_ted.tokenizer = tokenizer
model.smi_ted.set_padding_idx_from_tokenizer()

# Example SMILES strings
smiles = [
    'CC1C2CCC(C2)C1CN(CCO)C(=O)c1ccc(Cl)cc1',
    'COc1ccc(-c2cc(=O)c3c(O)c(OC)c(OC)cc3o2)cc1O',
    'CCOC(=O)c1ncn2c1CN(C)C(=O)c1cc(F)ccc1-2',
    'Clc1ccccc1-c1nc(-c2ccncc2)no1',
    'CC(C)(Oc1ccc(Cl)cc1)C(=O)OCc1cccc(CO)n1'
]

# Encode and decode SMILES
with torch.no_grad():
    encoder_outputs = model.encode(smiles)
    decoded_smiles = model.decode(encoder_outputs)

print(decoded_smiles)

📝 Citation

If you use this model, please cite:

@article{soares2025open,
  title={An open-source family of large encoder-decoder foundation models for chemistry},
  author={Soares, Eduardo and Vital Brazil, Emilio and Shirasuna, Victor and Zubarev, Dmitry and Cerqueira, Renato and Schmidt, Kristin},
  journal={Communications Chemistry},
  volume={8},
  number={1},
  pages={193},
  year={2025},
  publisher={Nature Publishing Group UK London}
}
@article{soares2024large,
  title={A large encoder-decoder family of foundation models for chemical language},
  author={Soares, Eduardo and Shirasuna, Victor and Brazil, Emilio Vital and Cerqueira, Renato and Zubarev, Dmitry and Schmidt, Kristin},
  journal={arXiv preprint arXiv:2407.20267},
  year={2024}
}

📧 Contact

For questions or collaborations, contact:

Note:
This fork adapts the original SMI-TED codebase for seamless integration with HuggingFace's AutoModel and AutoTokenizer interfaces. For full source code and training scripts, see the original IBM repo.